umer-wasim commited on
Commit
f1bdfc3
·
verified ·
1 Parent(s): f7e22d5

Security assistant chat as the front page (Nox-style, WebLLM on device, grounded in the Devseis knowledge base, Run audit from chat); phone check moved to phone.html

Browse files
README.md CHANGED
@@ -151,6 +151,19 @@ npm start # run the app
151
 
152
  Offline fine-tuning on CPU: `training/train_lora.py` (LoRA on Qwen2.5-0.5B-Instruct; see `--benchmark` and `--resume`).
153
 
 
 
 
 
 
 
 
 
 
 
 
 
 
154
  ## Install on a Mac
155
 
156
  Download from **https://devseis-endpoint-auditor.static.hf.space/download.html** (universal app for Apple silicon and
@@ -167,7 +180,7 @@ Code collects the evidence and decides every verdict; the fine-tuned model expla
167
 
168
  ## Phone check (iPhone, iPad, Android)
169
 
170
- The Space's front page (`index.html`) is a phone self-check that runs entirely in the browser:
171
 
172
  1. **Detects** what a browser can see: OS and version (Safari 26+ hides the iOS version in its user agent, so Safari's own version is used; Android uses client hints), device model (Android), browser, whether a device unlock method is set up (WebAuthn), and WebGPU support.
173
  2. **Asks** the rest, one question at a time with *where to look* in Settings (`catalog/mobile_questions.json`). Every result is labelled `detected` or `self-reported` in the evidence and the report.
 
151
 
152
  Offline fine-tuning on CPU: `training/train_lora.py` (LoRA on Qwen2.5-0.5B-Instruct; see `--benchmark` and `--resume`).
153
 
154
+ ## Security assistant (chat)
155
+
156
+ The Space's front page is a chat assistant in the style of Devseis Nox: ask about securing a computer or phone,
157
+ ISO 27001, GDPR or the EU AI Act, and run an audit from the chat. The model runs on the device with WebLLM (1.5B,
158
+ about 1 GB on first use); answers are grounded in the Devseis knowledge base (`library/knowledge.json`, 321 passages
159
+ generated by `library/build_knowledge.py` from the catalog and library), cite their sources, and are checked
160
+ (`chatProblems` in `app/renderer/knowledge.js`). Without WebGPU, or if the GPU fails, it answers directly from the
161
+ knowledge base. **Run audit**: on phones it starts the phone check (`phone.html`), which hands the report back to the
162
+ chat; on computers it leads to the desktop app, or the app's `…-evidence.json` report can be attached so the chat answers
163
+ from the user's own results and writes the full report. Conversations stay in the browser.
164
+ Until model v0.4 (Qwen2.5-1.5B fine-tuned for chat and reports) is published, the chat uses the base Qwen2.5-1.5B-Instruct
165
+ (`chat.version` in `app/model-version.json`).
166
+
167
  ## Install on a Mac
168
 
169
  Download from **https://devseis-endpoint-auditor.static.hf.space/download.html** (universal app for Apple silicon and
 
180
 
181
  ## Phone check (iPhone, iPad, Android)
182
 
183
+ The phone check (`phone.html`, also reachable from the chat's Run audit) runs entirely in the browser:
184
 
185
  1. **Detects** what a browser can see: OS and version (Safari 26+ hides the iOS version in its user agent, so Safari's own version is used; Android uses client hints), device model (Android), browser, whether a device unlock method is set up (WebAuthn), and WebGPU support.
186
  2. **Asks** the rest, one question at a time with *where to look* in Settings (`catalog/mobile_questions.json`). Every result is labelled `detected` or `self-reported` in the evidence and the report.
about.html CHANGED
@@ -54,7 +54,7 @@
54
  <main>
55
  <span class="tag">Open source · stage 1</span>
56
  <h1>Devseis Endpoint Auditor</h1>
57
- <p class="lead">A read-only audit of Windows, Linux, macOS, iPhone and Android devices against a security baseline, GDPR technical controls, EU AI Act controls and organisational controls, written up by a small AI auditor model that runs on the device. <a href="index.html">Check your phone now</a> or <a href="download.html">download the Mac app</a>.</p>
58
 
59
  <div class="grid">
60
  <div class="card"><b>Read-only collectors</b><span>PowerShell, Linux shell and macOS shell scripts. They read settings and never change the system.</span></div>
 
54
  <main>
55
  <span class="tag">Open source · stage 1</span>
56
  <h1>Devseis Endpoint Auditor</h1>
57
+ <p class="lead">A read-only audit of Windows, Linux, macOS, iPhone and Android devices against a security baseline, GDPR technical controls, EU AI Act controls and organisational controls, written up by a small AI auditor model that runs on the device. <a href="index.html">ask the assistant</a>, <a href="phone.html">check your phone</a> or <a href="download.html">download the Mac app</a>.</p>
58
 
59
  <div class="grid">
60
  <div class="card"><b>Read-only collectors</b><span>PowerShell, Linux shell and macOS shell scripts. They read settings and never change the system.</span></div>
app/model-version.json CHANGED
@@ -1,4 +1,9 @@
1
  {
2
  "version": "v0.3",
3
- "note": "Model version the desktop app and phone check use (Devseis/endpoint-auditor-0.5b-*-MLC tags). Read at start-up, so installed apps follow updates."
 
 
 
 
 
4
  }
 
1
  {
2
  "version": "v0.3",
3
+ "note": "Model version the desktop app and phone check use (Devseis/endpoint-auditor-0.5b-*-MLC tags). Read at start-up, so installed apps follow updates.",
4
+ "chat": {
5
+ "version": null,
6
+ "size": "1.5B",
7
+ "note": "Chat model: Devseis/endpoint-auditor-1.5b-*-MLC tag. null = base Qwen2.5-1.5B-Instruct until v0.4 is published."
8
+ }
9
  }
app/renderer/knowledge.js ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ // Chat grounding: search the Devseis knowledge base (library/knowledge.json) and build the chat prompt.
2
+ // Used by the web chat, the desktop app and (through training/chat-contexts.mjs) the training-data builder,
3
+ // so the model is trained on exactly the prompts it gets at runtime. Pure functions, no DOM or network.
4
+
5
+ export const CHAT_SYSTEM_PROMPT = 'You are the Devseis security assistant. You help people secure their computers and phones and '
6
+ + 'understand ISO 27001, GDPR and the EU AI Act. Answer in plain language that a non-technical person can follow, using only '
7
+ + 'the CONTEXT and MY AUDIT below, and cite the passages you use as [1], [2]. If the CONTEXT does not answer the question, say '
8
+ + 'you do not know rather than guessing. If the question is about the user\'s own device and MY AUDIT is "none", explain that '
9
+ + 'you cannot see their device and suggest running an audit. Never invent settings, versions or numbers.';
10
+
11
+ const STOP = new Set('a an and are as at be by can do does for from has have how i if in is it its me my of on or should so that the their them this to was what when where which who why will with you your'.split(' '));
12
+ // Words people use -> words the knowledge base uses.
13
+ const SYNONYMS = {
14
+ filevault: 'encryption disk', bitlocker: 'encryption disk', luks: 'encryption disk', encrypt: 'encryption', encrypted: 'encryption',
15
+ antivirus: 'malware antivirus', virus: 'malware', defender: 'malware antivirus defender', xprotect: 'malware', clamav: 'malware',
16
+ passcode: 'passcode screen lock', pin: 'passcode pin screen lock', autolock: 'auto lock screen timeout', lock: 'lock screen',
17
+ password: 'password', admin: 'administrator', administrator: 'administrator', sudo: 'administrator', root: 'administrator rooted',
18
+ jailbreak: 'jailbroken', jailbroken: 'jailbroken', rooted: 'rooted', update: 'updates update', updates: 'updates update', patch: 'updates security patch',
19
+ backup: 'backup', 'time machine': 'backup', icloud: 'icloud backup', usb: 'usb storage', 'usb stick': 'usb storage', thumb: 'usb storage',
20
+ firewall: 'firewall', vpn: 'remote access', rdp: 'remote access', teamviewer: 'remote access teamviewer', anydesk: 'remote access anydesk',
21
+ cve: 'vulnerabilities vulnerability', vulnerability: 'vulnerabilities vulnerability', vulnerable: 'vulnerabilities', ai: 'ai', chatgpt: 'chatgpt ai',
22
+ gpt: 'chatgpt ai', copilot: 'copilot ai', gemini: 'gemini ai', iphone: 'iphone ios', ipad: 'iphone ios ipados', mac: 'macos', macbook: 'macos',
23
+ osx: 'macos', windows: 'windows', win: 'windows', ubuntu: 'linux ubuntu', debian: 'linux debian', fedora: 'linux', android: 'android',
24
+ galaxy: 'android', pixel: 'android', samsung: 'android', gdpr: 'gdpr', iso: 'iso 27001', '27001': 'iso 27001', 'ai act': 'ai act',
25
+ install: 'install', download: 'install download', privacy: 'privacy', private: 'privacy', notifications: 'notification lock screen previews',
26
+ };
27
+ const OS_WORDS = { macos: ['macos', 'mac', 'macbook', 'osx', 'filevault'], windows: ['windows', 'win', 'bitlocker', 'defender'],
28
+ linux: ['linux', 'ubuntu', 'debian', 'fedora', 'luks'], ios: ['iphone', 'ipad', 'ios'], android: ['android', 'galaxy', 'pixel', 'samsung'] };
29
+
30
+ function stem(word) {
31
+ if (word.length > 5 && word.endsWith('ing')) return word.slice(0, -3);
32
+ if (word.length > 4 && word.endsWith('ed')) return word.slice(0, -2);
33
+ if (word.length > 4 && word.endsWith('es') && !word.endsWith('ses')) return word.slice(0, -2);
34
+ if (word.length > 3 && word.endsWith('s') && !word.endsWith('ss')) return word.slice(0, -1);
35
+ return word;
36
+ }
37
+ export function tokens(text, expand = false) {
38
+ let t = String(text).toLowerCase();
39
+ if (expand) for (const [k, v] of Object.entries(SYNONYMS)) if (k.includes(' ') && t.includes(k)) t += ' ' + v;
40
+ const words = t.split(/[^a-z0-9.]+/).map(w => w.replace(/^\.+|\.+$/g, '')).filter(w => w && !STOP.has(w));
41
+ const out = [];
42
+ for (const w of words) {
43
+ out.push(stem(w));
44
+ if (expand && SYNONYMS[w]) for (const s of SYNONYMS[w].split(' ')) out.push(stem(s));
45
+ }
46
+ return out;
47
+ }
48
+
49
+ // The operating system a question is about, if it names one.
50
+ export function osInQuestion(question) {
51
+ const words = new Set(String(question).toLowerCase().split(/[^a-z0-9]+/));
52
+ for (const [os, list] of Object.entries(OS_WORDS)) if (list.some(w => words.has(w))) return os;
53
+ return null;
54
+ }
55
+
56
+ export class Knowledge {
57
+ constructor(data) {
58
+ this.chunks = data.chunks;
59
+ this.docs = this.chunks.map(c => {
60
+ const tf = new Map();
61
+ for (const t of [...tokens(c.title), ...tokens(c.title), ...tokens(c.text)]) tf.set(t, (tf.get(t) || 0) + 1);
62
+ let len = 0; for (const v of tf.values()) len += v;
63
+ return { tf, len };
64
+ });
65
+ this.avg = this.docs.reduce((s, d) => s + d.len, 0) / this.docs.length;
66
+ this.df = new Map();
67
+ for (const d of this.docs) for (const t of d.tf.keys()) this.df.set(t, (this.df.get(t) || 0) + 1);
68
+ this.byId = new Map(this.chunks.map((c, i) => [c.id, i]));
69
+ }
70
+
71
+ // Top-k passages (BM25), preferring the OS the question names, else the device's OS.
72
+ search(question, { k = 4, os = null } = {}) {
73
+ const target = osInQuestion(question) || os;
74
+ const q = tokens(question, true);
75
+ const N = this.docs.length, k1 = 1.4, b = 0.75;
76
+ const scores = this.docs.map((d, i) => {
77
+ let s = 0;
78
+ for (const t of new Set(q)) {
79
+ const f = d.tf.get(t); if (!f) continue;
80
+ const idf = Math.log(1 + (N - this.df.get(t) + 0.5) / (this.df.get(t) + 0.5));
81
+ s += idf * (f * (k1 + 1)) / (f + k1 * (1 - b + b * d.len / this.avg));
82
+ }
83
+ const c = this.chunks[i];
84
+ if (target && c.os) s *= c.os === target ? 1.6 : 0.45;
85
+ return { i, s };
86
+ });
87
+ return scores.filter(x => x.s > 0).sort((a, b) => b.s - a.s || a.i - b.i).slice(0, k).map(x => this.chunks[x.i]);
88
+ }
89
+
90
+ get(id) { const i = this.byId.get(id); return i === undefined ? null : this.chunks[i]; }
91
+ }
92
+
93
+ // Short, model-friendly summary of an audit the user ran (evidence JSON), or "none".
94
+ export function auditBrief(evidence, maxGaps = 8) {
95
+ if (!evidence) return 'none';
96
+ const h = evidence.host, s = evidence.summary;
97
+ const assessed = s.Compliant + s['Non-Compliant'];
98
+ const lines = [`${h.os_name} ${h.os_version} (${h.os_family}), audited ${String(evidence.audit.started_at).slice(0, 10)}; `
99
+ + `${s.Compliant} of ${assessed} assessed checks compliant; ${s['Non-Compliant']} gaps, ${s.Pending} pending, ${s.Error} not read.`];
100
+ const gaps = evidence.results.filter(r => r.status === 'Non-Compliant').slice(0, maxGaps);
101
+ if (gaps.length) lines.push('Gaps: ' + gaps.map(r => `${r.name}: ${r.actual}`).join('; ') + '.');
102
+ return lines.join(' ');
103
+ }
104
+
105
+ // Messages for one chat turn. history: earlier [{role, content}] turns (already answered).
106
+ export function chatMessages(question, passages, audit = 'none', history = []) {
107
+ const context = passages.length ? passages.map((p, n) => `[${n + 1}] ${p.title}: ${p.text}`).join('\n') : '(no matching passages)';
108
+ const user = `CONTEXT:\n${context}\n\nMY AUDIT: ${audit}\n\nQUESTION: ${question}`;
109
+ return [{ role: 'system', content: CHAT_SYSTEM_PROMPT }, ...history.slice(-4), { role: 'user', content: user }];
110
+ }
111
+
112
+ // Grounding rules for a chat answer (same as check_chat() in training/build_chat.py).
113
+ const NUMBER = /(?<![\w.])\d+(?:\.\d+)?(?!\w)/g;
114
+ export function chatProblems(userPrompt, answer) {
115
+ const problems = [];
116
+ const passages = (userPrompt.split('MY AUDIT:')[0].match(/^\[\d+\] /gm) || []).length;
117
+ for (const m of answer.matchAll(/\[(\d+)\]/g)) if (+m[1] < 1 || +m[1] > passages) problems.push(`citation [${m[1]}] out of range`);
118
+ const allowed = new Set(userPrompt.match(NUMBER) || []);
119
+ for (const n of answer.match(NUMBER) || []) if (!allowed.has(n) && !(/^\d+$/.test(n) && +n <= 12)) problems.push(`number ${n} not in sources`);
120
+ return problems;
121
+ }
122
+
123
+ // With an audit attached: the fix guide for the gap the question names, else the top gap, is always among the
124
+ // passages (as in the training data), so "what should I fix first?" is answered from the right guide.
125
+ export function searchWithAudit(kb, question, evidence, { k = 4, os = null } = {}) {
126
+ let passages = kb.search(question, { k, os: evidence ? evidence.host.os_family : os });
127
+ if (!evidence) return passages;
128
+ const gaps = evidence.results.filter(r => r.status === 'Non-Compliant').slice(0, 8);
129
+ if (!gaps.length) return passages;
130
+ const q = String(question).toLowerCase();
131
+ const named = gaps.find(g => q.includes(g.name.toLowerCase()));
132
+ const gap = named || gaps[0];
133
+ const gold = kb.get(`fix:${gap.id}:${evidence.host.os_family}`);
134
+ if (gold && !passages.some(p => p.id === gold.id)) passages = [...passages.slice(0, k - 1), gold];
135
+ return passages;
136
+ }
chat/chat.css ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ /* Devseis Security Assistant: layout follows Devseis Nox (header, empty state with chips, thread, pill composer, sidebar). */
2
+ :root { --bg: #f4f6fb; --panel: #fff; --border: #e2e6ee; --text: #171a21; --muted: #666f80; --accent: #ED2939; --accent-dark: #C81E2B;
3
+ --accent-text: #fff; --accent-soft: rgba(237, 41, 57, .08); --accent-border: rgba(237, 41, 57, .35); --bubble-bot: #fff; --bubble-user: #ED2939;
4
+ --input-bg: #fff; --blue: #00A1DE; --ok: #1a7f37; --warn: #9a6700; }
5
+ @media (prefers-color-scheme: dark) { :root:not([data-theme="light"]) { --bg: #0b0e14; --panel: #11151d; --border: #232a37; --text: #e8ecf4; --muted: #8b94a7;
6
+ --accent: #ff4d5a; --accent-dark: #ED2939; --accent-soft: rgba(255, 77, 90, .14); --accent-border: rgba(255, 77, 90, .4); --bubble-bot: #161c27; --bubble-user: #c81e2b; --input-bg: #1a202b; --ok: #3fb950; --warn: #d29922; } }
7
+ :root[data-theme="dark"] { --bg: #0b0e14; --panel: #11151d; --border: #232a37; --text: #e8ecf4; --muted: #8b94a7;
8
+ --accent: #ff4d5a; --accent-dark: #ED2939; --accent-soft: rgba(255, 77, 90, .14); --accent-border: rgba(255, 77, 90, .4); --bubble-bot: #161c27; --bubble-user: #c81e2b; --input-bg: #1a202b; --ok: #3fb950; --warn: #d29922; }
9
+ * { box-sizing: border-box; }
10
+ html, body { height: 100%; }
11
+ body { margin: 0; background: var(--bg); color: var(--text); font: 16px/1.55 -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Arial, sans-serif;
12
+ height: 100dvh; display: flex; flex-direction: column; overflow: hidden; }
13
+ a { color: var(--blue); }
14
+ .hidden { display: none !important; }
15
+ .muted { color: var(--muted); } .tiny { font-size: 12px; }
16
+ button { font: inherit; cursor: pointer; }
17
+ button:focus-visible, a:focus-visible, .chip:focus-visible { outline: 2px solid var(--accent); outline-offset: 2px; }
18
+
19
+ header { display: flex; align-items: center; justify-content: space-between; gap: 10px; padding: 10px 14px; padding-top: calc(10px + env(safe-area-inset-top));
20
+ border-bottom: 1px solid var(--border); background: var(--panel); border-top: 3px solid var(--accent); flex-shrink: 0; }
21
+ .header-left, .header-actions { display: flex; align-items: center; gap: 6px; }
22
+ .brand { display: flex; align-items: center; gap: 8px; color: var(--text); text-decoration: none; font-size: 15px; }
23
+ .brand img { width: 28px; height: 28px; }
24
+ .icon-btn { width: 38px; height: 38px; border-radius: 10px; border: 1px solid transparent; background: transparent; color: var(--text); display: grid; place-items: center; }
25
+ .icon-btn:hover { background: var(--accent-soft); }
26
+ .audit-btn { background: var(--accent); color: var(--accent-text); border: 0; border-radius: 999px; padding: 8px 14px; font-weight: 700; font-size: 14px; white-space: nowrap; }
27
+ .audit-btn:hover { background: var(--accent-dark); }
28
+ .status-bar { padding: 6px 16px; background: var(--accent-soft); border-bottom: 1px solid var(--border); font-size: 13px; }
29
+ .status-progress { height: 4px; background: var(--border); border-radius: 99px; overflow: hidden; margin-top: 4px; }
30
+ #statusFill { height: 100%; width: 0; background: var(--accent); transition: width .3s; }
31
+
32
+ main { flex: 1; overflow-y: auto; padding: 18px 14px 8px; }
33
+ .empty-state { max-width: 680px; margin: 6vh auto 0; text-align: center; }
34
+ .hero-icon img { width: 64px; height: 64px; }
35
+ .empty-state h1 { font-size: clamp(24px, 4.5vw, 32px); line-height: 1.2; margin: 10px 0; }
36
+ .hero-sub { color: var(--muted); font-size: 17px; margin: 0 auto 20px; max-width: 560px; }
37
+ .chips { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); gap: 10px; max-width: 600px; margin: 0 auto; }
38
+ @media (max-width: 520px) { .chips { grid-template-columns: 1fr; } }
39
+ .chip { text-align: left; background: var(--panel); border: 1px solid var(--border); border-radius: 14px; padding: 12px 14px; color: var(--text); font-size: 15px; }
40
+ .chip:hover { border-color: var(--accent-border); background: var(--accent-soft); }
41
+
42
+ #thread { max-width: 780px; margin: 0 auto; display: flex; flex-direction: column; gap: 14px; }
43
+ .msg { display: flex; gap: 10px; align-items: flex-start; }
44
+ .msg.user { justify-content: flex-end; }
45
+ .avatar { width: 30px; height: 30px; flex: none; border-radius: 8px; }
46
+ .bubble { max-width: min(640px, 88%); padding: 11px 15px; border-radius: 18px; overflow-wrap: anywhere; }
47
+ .msg.user .bubble { background: var(--bubble-user); color: #fff; border-bottom-right-radius: 6px; white-space: pre-wrap; }
48
+ .msg.bot .bubble { background: var(--bubble-bot); border: 1px solid var(--border); border-bottom-left-radius: 6px; }
49
+ .bubble p { margin: 0 0 8px; } .bubble p:last-child { margin-bottom: 0; }
50
+ .bubble ol, .bubble ul { margin: 6px 0 8px 20px; padding: 0; } .bubble li { margin: 3px 0; }
51
+ .bubble code { font: 13px ui-monospace, Menlo, monospace; background: var(--accent-soft); padding: 1px 5px; border-radius: 5px; }
52
+ .cite { display: inline-block; font-size: 11px; font-weight: 700; color: var(--accent); background: var(--accent-soft); border-radius: 6px; padding: 0 5px; margin: 0 1px; cursor: pointer; vertical-align: 1px; border: 0; }
53
+ .sources { margin-top: 10px; display: flex; flex-wrap: wrap; gap: 6px; }
54
+ .source { font-size: 12px; color: var(--muted); border: 1px solid var(--border); border-radius: 999px; padding: 2px 9px; background: transparent; }
55
+ .source.open { border-color: var(--accent-border); color: var(--text); }
56
+ .source-text { font-size: 13px; color: var(--muted); border-left: 3px solid var(--accent-border); padding: 4px 10px; margin-top: 8px; }
57
+ .note { font-size: 12px; color: var(--warn); margin-top: 8px; }
58
+ .typing span { display: inline-block; width: 7px; height: 7px; border-radius: 50%; background: var(--muted); margin-right: 3px; animation: blink 1.2s infinite; }
59
+ .typing span:nth-child(2) { animation-delay: .2s; } .typing span:nth-child(3) { animation-delay: .4s; }
60
+ @keyframes blink { 0%, 80%, 100% { opacity: .25; } 40% { opacity: 1; } }
61
+ .card { border: 1px solid var(--border); background: var(--panel); border-radius: 16px; padding: 14px; }
62
+ .card .score { font-size: 30px; font-weight: 800; }
63
+ .card .pill { display: inline-block; font-size: 12px; font-weight: 700; border: 1px solid; border-radius: 99px; padding: 0 8px; margin: 2px 4px 2px 0; }
64
+ .btn-row { display: flex; flex-wrap: wrap; gap: 8px; margin-top: 10px; }
65
+ .small-btn { border: 1px solid var(--border); background: var(--panel); color: var(--text); border-radius: 10px; padding: 7px 12px; font-size: 14px; }
66
+ .primary-btn { border: 0; background: var(--accent); color: #fff; border-radius: 10px; padding: 8px 14px; font-size: 14px; font-weight: 700; }
67
+
68
+ #composer { display: flex; align-items: flex-end; gap: 8px; max-width: 812px; width: 100%; margin: 0 auto; padding: 8px 14px 4px; flex-shrink: 0; }
69
+ .input-pill { flex: 1; display: flex; background: var(--input-bg); border: 1px solid var(--border); border-radius: 22px; padding: 4px 6px 4px 16px; }
70
+ .input-pill:focus-within { box-shadow: 0 0 0 2px var(--accent-border); }
71
+ #input { flex: 1; border: 0; outline: 0; resize: none; background: transparent; color: var(--text); font: inherit; max-height: 180px; padding: 9px 0; }
72
+ .send-btn { width: 44px; height: 44px; border-radius: 50%; border: 0; background: var(--accent); color: #fff; display: grid; place-items: center; flex: none; }
73
+ .send-btn:disabled { opacity: .4; cursor: default; }
74
+ .attach { width: 44px; height: 44px; flex: none; }
75
+ .disclaimer { text-align: center; font-size: 11.5px; color: var(--muted); margin: 0 0 calc(8px + env(safe-area-inset-bottom)); padding: 0 12px; }
76
+
77
+ .overlay { position: fixed; inset: 0; background: rgba(0,0,0,.35); z-index: 9; }
78
+ .sidebar { position: fixed; top: 0; bottom: 0; left: 0; width: min(320px, 86vw); background: var(--panel); border-right: 1px solid var(--border); z-index: 10;
79
+ transform: translateX(-102%); transition: transform .2s; display: flex; flex-direction: column; padding-top: env(safe-area-inset-top); }
80
+ .sidebar.open { transform: none; }
81
+ .sidebar-head { display: flex; justify-content: space-between; align-items: center; padding: 12px 14px; border-bottom: 1px solid var(--border); }
82
+ .sidebar-head h2 { font-size: 16px; margin: 0; }
83
+ .sidebar-list { flex: 1; overflow-y: auto; padding: 8px; }
84
+ .convo { display: flex; align-items: center; gap: 6px; padding: 9px 10px; border-radius: 10px; cursor: pointer; font-size: 14px; }
85
+ .convo:hover, .convo.active { background: var(--accent-soft); }
86
+ .convo span { flex: 1; white-space: nowrap; overflow: hidden; text-overflow: ellipsis; }
87
+ .convo button { opacity: .5; border: 0; background: transparent; color: var(--text); }
88
+ .sidebar-foot { border-top: 1px solid var(--border); padding: 12px 14px calc(12px + env(safe-area-inset-bottom)); font-size: 14px; }
89
+ .sidebar-foot .row { display: flex; justify-content: space-between; align-items: center; gap: 8px; margin-bottom: 8px; }
90
+
91
+ .modal { position: fixed; inset: 0; background: rgba(0,0,0,.45); display: grid; place-items: center; z-index: 20; padding: 16px; }
92
+ .modal-card { background: var(--panel); border-radius: 18px; padding: 20px; max-width: 460px; width: 100%; border: 1px solid var(--border); }
93
+ .modal-card h2 { margin: 0 0 6px; font-size: 20px; }
94
+ .option { display: block; width: 100%; text-align: left; border: 1px solid var(--border); background: var(--bg); color: var(--text); border-radius: 14px; padding: 12px 14px; margin: 10px 0; text-decoration: none; }
95
+ .option:hover { border-color: var(--accent-border); }
96
+ .option b { display: block; } .option span { color: var(--muted); font-size: 14px; }
97
+ .option.recommended { border: 2px solid var(--accent); }
98
+ .link-btn { border: 0; background: transparent; color: var(--muted); text-decoration: underline; padding: 6px 0; }
99
+ @media (max-width: 560px) { .brand span { display: none; } .audit-btn { padding: 8px 12px; } .hero-sub { font-size: 16px; } }
chat/chat.js ADDED
@@ -0,0 +1,310 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ // Devseis Security Assistant: chat about endpoint security with the Devseis model running on the device (WebLLM),
2
+ // grounded in the Devseis knowledge base (library/knowledge.json), plus audits: the phone check on phones, the desktop
3
+ // app (or its report file) on computers. Conversations are stored only in this browser.
4
+ import { Knowledge, chatMessages, auditBrief, chatProblems, searchWithAudit } from '../app/renderer/knowledge.js';
5
+ import { Library, findingPrompt, findingTemplate, summaryPrompt, summaryTemplate, aiPrompt, aiTemplate, writeAnswer, ANSWER_SCHEMAS } from '../app/renderer/auditor-core.js';
6
+ import { buildReportHtml } from '../app/renderer/report.js';
7
+ import { detectDevice } from '../mobile/detect.js';
8
+
9
+ const WEBLLM_URL = 'https://esm.run/@mlc-ai/web-llm@0.2.85';
10
+ const STORE = 'devseis.chats.v1';
11
+ const $ = id => document.getElementById(id);
12
+ const esc = s => String(s ?? '').replace(/[&<>"']/g, c => ({ '&': '&amp;', '<': '&lt;', '>': '&gt;', '"': '&quot;', "'": '&#39;' }[c]));
13
+
14
+ const state = { kb: null, device: null, config: null, engine: null, enginePromise: null, modelLabel: '', busy: false, chats: [], current: null, libData: null };
15
+
16
+ // ---------- storage ----------
17
+ function loadChats() { try { return JSON.parse(localStorage.getItem(STORE)) || []; } catch { return []; } }
18
+ function saveChats() { try { localStorage.setItem(STORE, JSON.stringify(state.chats.slice(0, 50))); } catch { /* private mode or full */ } }
19
+ function newChat() {
20
+ const chat = { id: Date.now().toString(36), title: 'New conversation', messages: [], audit: null };
21
+ state.chats.unshift(chat); state.current = chat; saveChats(); render(); return chat;
22
+ }
23
+
24
+ // ---------- model ----------
25
+ async function getEngine() {
26
+ if (state.engine !== null || state.enginePromise) return state.enginePromise || state.engine;
27
+ state.enginePromise = (async () => {
28
+ if (!('gpu' in navigator)) throw new Error('This browser has no WebGPU');
29
+ const adapter = await navigator.gpu.requestAdapter();
30
+ if (!adapter) throw new Error('No WebGPU graphics adapter');
31
+ const quant = adapter.features.has('shader-f16') ? 'q4f16_1' : 'q4f32_1';
32
+ const webllm = await import(WEBLLM_URL);
33
+ const base = webllm.prebuiltAppConfig.model_list.find(m => m.model_id === `Qwen2.5-1.5B-Instruct-${quant}-MLC`);
34
+ const v = state.config?.chat?.version;
35
+ const custom = v ? { ...base, model: `https://huggingface.co/Devseis/endpoint-auditor-1.5b-${quant}-MLC/resolve/${v}/`, model_id: `devseis-assistant-1.5b-${v}-${quant}` } : null;
36
+ const appConfig = { ...webllm.prebuiltAppConfig, model_list: custom ? [...webllm.prebuiltAppConfig.model_list, custom] : webllm.prebuiltAppConfig.model_list };
37
+ const initProgressCallback = r => status(`Getting the AI ready on this device (first time only, about 1 GB): ${Math.round((r.progress || 0) * 100)}%`, r.progress || 0);
38
+ let engine, label;
39
+ if (custom) {
40
+ try { engine = await webllm.CreateMLCEngine(custom.model_id, { appConfig, initProgressCallback }); label = `Devseis assistant ${v} (1.5B)`; } catch (e) { console.warn('Devseis model failed, using base', e); }
41
+ }
42
+ if (!engine) { engine = await webllm.CreateMLCEngine(base.model_id, { appConfig, initProgressCallback }); label = 'Qwen2.5 1.5B (base, until the Devseis model loads)'; }
43
+ state.modelLabel = label; $('modelName').textContent = label; status('');
44
+ return engine;
45
+ })().then(e => (state.engine = e), err => { state.engine = false; state.modelLabel = `Not available: ${err.message}`; $('modelName').textContent = 'Knowledge base only'; status(''); return false; });
46
+ return state.enginePromise;
47
+ }
48
+
49
+ function status(text, fraction) {
50
+ $('statusBar').classList.toggle('hidden', !text);
51
+ $('statusText').textContent = text;
52
+ $('statusFill').style.width = `${Math.round((fraction || 0) * 100)}%`;
53
+ }
54
+
55
+ // ---------- rendering ----------
56
+ function markdown(text, passages, msgIndex) {
57
+ const lines = esc(text).split('\n');
58
+ let html = '', list = null;
59
+ const close = () => { if (list) { html += `</${list}>`; list = null; } };
60
+ for (const line of lines) {
61
+ const ol = /^\s*\d+\.\s+(.*)$/.exec(line), ul = /^\s*[-•]\s+(.*)$/.exec(line);
62
+ if (ol || ul) {
63
+ const tag = ol ? 'ol' : 'ul';
64
+ if (list !== tag) { close(); html += `<${tag}>`; list = tag; }
65
+ html += `<li>${(ol || ul)[1]}</li>`;
66
+ } else { close(); if (line.trim()) html += `<p>${line}</p>`; }
67
+ }
68
+ close();
69
+ return html.replace(/\*\*(.+?)\*\*/g, '<b>$1</b>').replace(/`([^`]+)`/g, '<code>$1</code>')
70
+ .replace(/\[(\d+)\]/g, (m, n) => passages && passages[n - 1] ? `<button type="button" class="cite" data-msg="${msgIndex}" data-n="${n}" title="${esc(passages[n - 1].title)}">${n}</button>` : '');
71
+ }
72
+
73
+ function auditCard(audit) {
74
+ const s = audit.summary, assessed = s.Compliant + s['Non-Compliant'];
75
+ const pct = assessed ? Math.round(100 * s.Compliant / assessed) : 0;
76
+ const colours = { Compliant: 'var(--ok)', 'Non-Compliant': 'var(--accent)', Pending: 'var(--warn)', NotApplicable: 'var(--muted)', Error: '#8250df' };
77
+ const gaps = audit.results.filter(r => r.status === 'Non-Compliant').slice(0, 5);
78
+ return `<div class="card"><div class="muted tiny">AUDIT REPORT · ${esc(audit.host.os_name)} ${esc(audit.host.os_version)} · ${esc(String(audit.audit.started_at).slice(0, 10))}</div>
79
+ <div class="score">${pct}%</div><div class="muted">${s.Compliant} of ${assessed} assessed checks compliant</div>
80
+ <div>${Object.entries(s).filter(([, v]) => v).map(([k, v]) => `<span class="pill" style="color:${colours[k]}">${v} ${k}</span>`).join('')}</div>
81
+ ${gaps.length ? `<p style="margin:8px 0 0"><b>Top gaps:</b> ${gaps.map(g => esc(g.name)).join(', ')}</p>` : ''}
82
+ <div class="btn-row"><button type="button" class="primary-btn" data-ask="What should I fix first?">What should I fix first?</button>
83
+ <button type="button" class="small-btn" data-report="1">📄 Write my full report</button></div></div>`;
84
+ }
85
+
86
+ function render() {
87
+ const chat = state.current;
88
+ $('emptyState').classList.toggle('hidden', Boolean(chat && (chat.messages.length || chat.audit)));
89
+ const thread = $('thread');
90
+ thread.innerHTML = '';
91
+ if (!chat) return renderSidebar();
92
+ if (chat.audit) thread.insertAdjacentHTML('beforeend', `<div class="msg bot"><img class="avatar" src="app/build/icon.png" alt="">${auditCard(chat.audit)}</div>`);
93
+ chat.messages.forEach((m, i) => {
94
+ if (m.role === 'user') { thread.insertAdjacentHTML('beforeend', `<div class="msg user"><div class="bubble">${esc(m.content)}</div></div>`); return; }
95
+ const sources = (m.passages || []).map((p, n) => `<button type="button" class="source" data-msg="${i}" data-n="${n + 1}">[${n + 1}] ${esc(p.title)}</button>`).join('');
96
+ thread.insertAdjacentHTML('beforeend', `<div class="msg bot"><img class="avatar" src="app/build/icon.png" alt=""><div class="bubble" id="m${i}">
97
+ ${m.pending ? '<div class="typing"><span></span><span></span><span></span></div>' : markdown(m.content, m.passages, i)}
98
+ ${!m.pending && sources && !m.noSources ? `<div class="sources"><span class="muted tiny">Sources:</span>${sources}</div>` : ''}
99
+ ${m.note ? `<div class="note">${esc(m.note)}</div>` : ''}<div class="source-slot"></div></div></div>`);
100
+ });
101
+ renderSidebar();
102
+ $('chatMain').scrollTop = $('chatMain').scrollHeight;
103
+ }
104
+
105
+ function renderSidebar() {
106
+ $('sidebarList').innerHTML = state.chats.map(c => `<div class="convo${c === state.current ? ' active' : ''}" data-id="${c.id}"><span>${esc(c.title)}</span><button type="button" data-del="${c.id}" aria-label="Delete">✕</button></div>`).join('')
107
+ || '<p class="muted tiny" style="padding:8px">No conversations yet.</p>';
108
+ }
109
+
110
+ function showSource(msgIndex, n, anchor) {
111
+ const m = state.current.messages[msgIndex];
112
+ const p = m && m.passages && m.passages[n - 1];
113
+ if (!p) return;
114
+ const slot = document.querySelector(`#m${msgIndex} .source-slot`);
115
+ slot.innerHTML = `<div class="source-text"><b>[${n}] ${esc(p.title)}</b><br>${esc(p.text)}</div>`;
116
+ document.querySelectorAll(`#m${msgIndex} .source`).forEach(b => b.classList.toggle('open', b.dataset.n === String(n)));
117
+ if (anchor) slot.scrollIntoView({ block: 'nearest' });
118
+ }
119
+
120
+ // ---------- asking ----------
121
+ async function ask(question) {
122
+ if (state.busy || !question.trim()) return;
123
+ const chat = state.current || newChat();
124
+ if (chat.title === 'New conversation') chat.title = question.slice(0, 60);
125
+ const history = chat.messages.filter(m => !m.pending).slice(-4).map(m => ({ role: m.role, content: m.content }));
126
+ chat.messages.push({ role: 'user', content: question });
127
+ const reply = { role: 'assistant', content: '', pending: true, passages: [] };
128
+ chat.messages.push(reply);
129
+ state.busy = true; updateSend(); render();
130
+ try {
131
+ const os = chat.audit ? chat.audit.host.os_family : (state.device && (state.device.os_family || (state.device.desktop ? desktopOs() : null)));
132
+ const passages = searchWithAudit(state.kb, question, chat.audit, { k: 4, os });
133
+ reply.passages = passages;
134
+ const messages = chatMessages(question, passages, auditBrief(chat.audit), history);
135
+ let engine = await getEngine();
136
+ let streamed = false;
137
+ if (engine) {
138
+ // A GPU can be lost (memory pressure, sleep); restart the model once, then fall back to the knowledge base.
139
+ for (let attempt = 0; attempt < 2 && engine; attempt++) {
140
+ try {
141
+ reply.pending = false; reply.content = '';
142
+ const stream = await engine.chat.completions.create({ messages, temperature: 0, max_tokens: 700, stream: true });
143
+ let last = 0;
144
+ for await (const chunk of stream) {
145
+ reply.content += chunk.choices[0]?.delta?.content || '';
146
+ if (Date.now() - last > 120) { last = Date.now(); const el = $(`m${chat.messages.length - 1}`); if (el) el.innerHTML = markdown(clean(reply.content), passages, chat.messages.length - 1); }
147
+ }
148
+ streamed = true;
149
+ break;
150
+ } catch (error) {
151
+ if (!/lost|Instance reference|GPUDevice|disposed/i.test(error.message)) throw error;
152
+ status('The graphics processor was reset; restarting the AI…', 0.5);
153
+ state.engine = null; state.enginePromise = null;
154
+ engine = attempt === 0 ? await getEngine() : null;
155
+ }
156
+ }
157
+ }
158
+ if (streamed) {
159
+ reply.content = clean(reply.content);
160
+ reply.noSources = !/\[\d+\]/.test(reply.content) && /can only help|don't have|do not have|not able to|can't answer|cannot answer|sorry/i.test(reply.content);
161
+ const problems = chatProblems(messages[messages.length - 1].content, reply.content);
162
+ reply.cited = /\[\d+\]/.test(reply.content);
163
+ if (problems.length) reply.note = 'Some details in this answer could not be matched to the sources. Check them in the sources below before acting.';
164
+ } else {
165
+ reply.content = passages.length
166
+ ? `The AI model isn't available right now, but here is what the Devseis knowledge base says:\n\n${passages.slice(0, 2).map((p, i) => `**${p.title}** [${i + 1}]\n${p.text}`).join('\n\n')}`
167
+ : 'The AI model isn\'t available right now, and the Devseis knowledge base has nothing on that question. I cover device security, ISO 27001, GDPR and the EU AI Act.';
168
+ reply.cited = true; reply.note = 'Shown directly from the knowledge base (the AI model is not available on this device right now).';
169
+ }
170
+ } catch (error) {
171
+ reply.content = `Something went wrong: ${error.message}`;
172
+ }
173
+ reply.pending = false; state.busy = false; saveChats(); updateSend(); render();
174
+ }
175
+
176
+ // Models sometimes emit HTML entities; show plain characters.
177
+ function clean(text) { return text.replace(/&amp;/g, '&').replace(/&lt;/g, '<').replace(/&gt;/g, '>').replace(/&quot;/g, '"').replace(/&#39;/g, "'"); }
178
+
179
+ function desktopOs() { const ua = navigator.userAgent; return /Windows/.test(ua) ? 'windows' : /Mac/.test(ua) ? 'macos' : 'linux'; }
180
+
181
+ // ---------- audits ----------
182
+ function openAuditModal() {
183
+ const d = state.device || {};
184
+ const phone = d.os_family === 'ios' || d.os_family === 'android';
185
+ const os = phone ? null : desktopOs();
186
+ const osName = { macos: 'Mac', windows: 'Windows', linux: 'Linux' }[os];
187
+ $('auditOptions').innerHTML = phone
188
+ ? `<p class="muted">The check runs in this browser: the phone answers what it can, you answer the rest. About 5 minutes.</p>
189
+ <a class="option recommended" href="phone.html?return=chat"><b>📱 Start the phone check</b><span>Your answers and report stay on this phone.</span></a>`
190
+ : `<p class="muted">A web page can't read a computer's settings, so computers are audited with the free Devseis app. It only reads, changes nothing, and takes about 15 minutes.</p>
191
+ <a class="option recommended" href="download.html#${os}"><b>💻 Get the desktop app for ${osName}${os === 'windows' ? ' (beta)' : ''}</b><span>Step-by-step install guide.</span></a>
192
+ <button type="button" class="option" id="optAttach"><b>📎 I already ran an audit</b><span>Add the report file (<code>…-evidence.json</code>, saved next to the PDF) and I'll answer from your results.</span></button>
193
+ <a class="option" href="phone.html"><b>📱 Check a phone instead</b><span>Open the phone check on the phone itself.</span></a>`;
194
+ $('auditModal').classList.remove('hidden');
195
+ const att = $('optAttach'); if (att) att.addEventListener('click', () => { $('auditModal').classList.add('hidden'); $('fileInput').click(); });
196
+ }
197
+
198
+ function attachAudit(evidence) {
199
+ if (!evidence || evidence.schema_version !== '1.0' || !Array.isArray(evidence.results) || !evidence.host) throw new Error('This is not a Devseis audit report file.');
200
+ const chat = state.current && !state.current.messages.length && !state.current.audit ? state.current : newChat();
201
+ chat.audit = evidence; chat.title = `Audit: ${evidence.host.os_name} ${String(evidence.audit.started_at).slice(0, 10)}`;
202
+ saveChats(); render();
203
+ }
204
+
205
+ async function loadLibrary() {
206
+ if (state.libData) return state.libData;
207
+ const get = async p => { const r = await fetch(p); if (!r.ok) throw new Error(`${p}: ${r.status}`); return r.json(); };
208
+ const [catalog, controls, guidance, aiTools] = await Promise.all(['catalog/checks.json', 'library/controls.json', 'library/check_guidance.json', 'library/ai_tools.json'].map(get));
209
+ state.libData = new Library({ catalog, controls, guidance, aiTools });
210
+ return state.libData;
211
+ }
212
+
213
+ async function writeReport() {
214
+ const chat = state.current; if (!chat || !chat.audit || state.busy) return;
215
+ state.busy = true; updateSend();
216
+ const evidence = chat.audit;
217
+ try {
218
+ const lib = await loadLibrary();
219
+ const engine = await getEngine();
220
+ const generate = engine ? async (system, prompt, task) => (await engine.chat.completions.create({
221
+ messages: [{ role: 'system', content: system }, { role: 'user', content: prompt }], temperature: 0, max_tokens: task === 'ai_classification' ? 1600 : 1200,
222
+ response_format: { type: 'json_object', schema: JSON.stringify(ANSWER_SCHEMAS[task]) } })).choices[0].message.content : null;
223
+ const targets = evidence.results.filter(r => ['Non-Compliant', 'Error', 'Pending'].includes(r.status));
224
+ const findings = {}, sources = { model: 0, template: 0 };
225
+ let i = 0;
226
+ for (const r of targets) {
227
+ status(`Writing your report: ${r.name} (${++i} of ${targets.length})`, i / (targets.length + 2));
228
+ const out = await writeAnswer('finding', findingPrompt(evidence, r, lib), findingTemplate(evidence, r, lib), r.section === 'org' ? null : generate);
229
+ findings[r.id] = out.answer; sources[out.source] += 1;
230
+ }
231
+ status('Writing your report: summary', 0.95);
232
+ const base = summaryTemplate(evidence, lib);
233
+ const so = await writeAnswer('summary', summaryPrompt(evidence, lib), base, generate);
234
+ sources[so.source] += 1;
235
+ const summary = so.source === 'model' ? { ...base, summary: so.answer.summary, next_steps: so.answer.next_steps } : base;
236
+ const ai = aiPrompt(evidence, lib);
237
+ let aiResult = null;
238
+ if (ai) { const o = await writeAnswer('ai_classification', ai, aiTemplate(evidence, lib), generate); aiResult = o.answer; sources[o.source] += 1; }
239
+ const html = buildReportHtml({ evidence, findings, summary, ai: aiResult, sha256: '(from attached file)', model: state.engine ? state.modelLabel : null, sources });
240
+ const url = URL.createObjectURL(new Blob([html], { type: 'text/html' }));
241
+ chat.messages.push({ role: 'assistant', content: `Your full report is ready: ${sources.model} parts written by the AI, ${sources.template} by the built-in writer. Open it and use your browser's Print › Save as PDF to keep it.`, passages: [] });
242
+ saveChats(); render();
243
+ const last = $(`m${chat.messages.length - 1}`);
244
+ last.insertAdjacentHTML('beforeend', `<div class="btn-row"><a class="primary-btn" href="${url}" target="_blank" rel="noopener" style="text-decoration:none">📄 Open the report</a></div>`);
245
+ } catch (error) {
246
+ chat.messages.push({ role: 'assistant', content: `The report could not be written: ${error.message}`, passages: [] }); render();
247
+ }
248
+ status(''); state.busy = false; updateSend();
249
+ }
250
+
251
+ // ---------- wiring ----------
252
+ function updateSend() { $('sendBtn').disabled = state.busy || !$('input').value.trim(); }
253
+ function closeSidebar() { $('sidebar').classList.remove('open'); $('overlay').classList.add('hidden'); }
254
+
255
+ $('composer').addEventListener('submit', e => { e.preventDefault(); const q = $('input').value.trim(); $('input').value = ''; $('input').style.height = ''; updateSend(); ask(q); });
256
+ $('input').addEventListener('input', () => { const t = $('input'); t.style.height = 'auto'; t.style.height = `${Math.min(t.scrollHeight, 180)}px`; updateSend(); });
257
+ $('input').addEventListener('keydown', e => { if (e.key === 'Enter' && !e.shiftKey && !e.isComposing) { e.preventDefault(); $('composer').requestSubmit(); } });
258
+ document.querySelectorAll('.chip').forEach(c => c.addEventListener('click', () => (c.dataset.action === 'audit' ? openAuditModal() : ask(c.dataset.prompt))));
259
+ $('auditBtn').addEventListener('click', openAuditModal);
260
+ $('auditClose').addEventListener('click', () => $('auditModal').classList.add('hidden'));
261
+ $('auditModal').addEventListener('click', e => { if (e.target === $('auditModal')) $('auditModal').classList.add('hidden'); });
262
+ $('attachBtn').addEventListener('click', () => $('fileInput').click());
263
+ $('fileInput').addEventListener('change', async () => {
264
+ const f = $('fileInput').files[0]; $('fileInput').value = '';
265
+ if (!f) return;
266
+ try { attachAudit(JSON.parse((await f.text()).replace(/^/, ''))); } catch (error) {
267
+ const chat = state.current || newChat(); chat.messages.push({ role: 'assistant', content: `I couldn't read that file: ${error.message} Use the …-evidence.json file the Devseis app saves next to the PDF report.`, passages: [] }); render();
268
+ }
269
+ });
270
+ $('thread').addEventListener('click', e => {
271
+ const t = e.target.closest('button'); if (!t) return;
272
+ if (t.dataset.ask) ask(t.dataset.ask);
273
+ else if (t.dataset.report) writeReport();
274
+ else if (t.dataset.msg) showSource(+t.dataset.msg, +t.dataset.n, true);
275
+ });
276
+ $('newChat').addEventListener('click', () => { newChat(); $('input').focus(); });
277
+ $('sidebarToggle').addEventListener('click', () => { $('sidebar').classList.add('open'); $('overlay').classList.remove('hidden'); });
278
+ $('sidebarClose').addEventListener('click', closeSidebar);
279
+ $('overlay').addEventListener('click', closeSidebar);
280
+ $('sidebarList').addEventListener('click', e => {
281
+ const del = e.target.closest('[data-del]');
282
+ if (del) { state.chats = state.chats.filter(c => c.id !== del.dataset.del); if (state.current && state.current.id === del.dataset.del) state.current = state.chats[0] || null; saveChats(); render(); return; }
283
+ const row = e.target.closest('.convo'); if (row) { state.current = state.chats.find(c => c.id === row.dataset.id); render(); closeSidebar(); }
284
+ });
285
+ $('clearAll').addEventListener('click', () => { state.chats = []; state.current = null; saveChats(); render(); });
286
+ $('themeToggle').addEventListener('click', () => {
287
+ const dark = document.documentElement.dataset.theme ? document.documentElement.dataset.theme === 'dark' : matchMedia('(prefers-color-scheme: dark)').matches;
288
+ document.documentElement.dataset.theme = dark ? 'light' : 'dark';
289
+ try { localStorage.setItem('devseis.theme', document.documentElement.dataset.theme); } catch { /* ignore */ }
290
+ });
291
+
292
+ async function init() {
293
+ try { const t = localStorage.getItem('devseis.theme'); if (t) document.documentElement.dataset.theme = t; } catch { /* ignore */ }
294
+ state.chats = loadChats();
295
+ state.current = state.chats[0] || null;
296
+ const [kb, device, config] = await Promise.all([
297
+ fetch('library/knowledge.json').then(r => r.json()),
298
+ detectDevice().catch(() => ({})),
299
+ fetch('app/model-version.json', { cache: 'no-cache' }).then(r => r.json()).catch(() => ({})),
300
+ ]);
301
+ state.kb = new Knowledge(kb); state.device = device; state.config = config;
302
+ // Coming back from the phone check with a fresh report.
303
+ if (location.hash === '#audit') {
304
+ try { const ev = JSON.parse(localStorage.getItem('devseis.lastAudit')); if (ev) { attachAudit(ev); localStorage.removeItem('devseis.lastAudit'); } } catch { /* ignore */ }
305
+ history.replaceState(null, '', location.pathname);
306
+ }
307
+ render();
308
+ if (new URLSearchParams(location.search).has('test')) Object.assign(window, { __chat: state, __ask: ask });
309
+ }
310
+ init();
docs/MODEL_LLM.md CHANGED
@@ -107,6 +107,8 @@ date it was built with, so a dataset version stays reproducible.
107
  | **auditor-0.5b v0.3** | v0.3 (tag `v0.3`) | 3,000 balanced, including phones (out of 39,500; none over 3,072 tokens) | Windows, Linux, macOS, iPhone, Android; 74 checks; real CVEs; evidence labels; computed summary scores | **In use** by the desktop apps and the phone check since 2026-10-10. Trained 375 steps (final validation loss 0.0109). Held-out: 100% valid, 100% grounded, 100% exact fields (60 examples). WebLLM on the GPU: q0f16 8/8 and q4f16_1 8/8 grounded on phone cases. Published with tag `v0.3` in all four model repositories |
108
  | auditor-0.5b v0.4+ | v0.3 + expert-reviewed library + gold findings | more, on a GPU if available | same, better wording and priorities | Later |
109
 
 
 
110
  Dataset v0.2 is not trained on: it was superseded by v0.3 before a run started.
111
 
112
  **Settings (all runs):** Qwen2.5-0.5B-Instruct, LoRA r=16, alpha=32 (8.8 M trainable parameters, 1.75%), learning rate
 
107
  | **auditor-0.5b v0.3** | v0.3 (tag `v0.3`) | 3,000 balanced, including phones (out of 39,500; none over 3,072 tokens) | Windows, Linux, macOS, iPhone, Android; 74 checks; real CVEs; evidence labels; computed summary scores | **In use** by the desktop apps and the phone check since 2026-10-10. Trained 375 steps (final validation loss 0.0109). Held-out: 100% valid, 100% grounded, 100% exact fields (60 examples). WebLLM on the GPU: q0f16 8/8 and q4f16_1 8/8 grounded on phone cases. Published with tag `v0.3` in all four model repositories |
108
  | auditor-0.5b v0.4+ | v0.3 + expert-reviewed library + gold findings | more, on a GPU if available | same, better wording and priorities | Later |
109
 
110
+ | auditor-1.5b v0.4 | v0.4 = v0.3 report tasks + 1,286 grounded chat examples (fixes, why, risk, controls, AI tools, phone settings, statuses, the auditor, the user's own audit, no-audit, out-of-scope, unknown) | 2,400 balanced (about 60% report, 40% chat) | Qwen2.5-**1.5B**; chat **and** reports in one model | **Training** since 2026-10-10 12:00 (about 300 steps, about 30 hours). Will power the chat and replace 0.5B v0.3 |
111
+
112
  Dataset v0.2 is not trained on: it was superseded by v0.3 before a run started.
113
 
114
  **Settings (all runs):** Qwen2.5-0.5B-Instruct, LoRA r=16, alpha=32 (8.8 M trainable parameters, 1.75%), learning rate
download.html CHANGED
@@ -69,7 +69,7 @@
69
  <body>
70
  <header class="top">
71
  <div class="brand"><img src="app/build/icon.png" alt=""><span>Devseis <b>Endpoint Auditor</b></span></div>
72
- <nav><a href="index.html">Phone check</a><a href="about.html">About</a></nav>
73
  </header>
74
 
75
  <main>
@@ -170,10 +170,10 @@
170
  <details><summary>What is in the report?</summary><p>A score, the most important gaps, and for each problem: what was found, why it matters, how risky it is, and simple steps to fix it — mapped to ISO 27001, GDPR and the EU AI Act.</p></details>
171
  <details><summary>Do I need internet?</summary><p>Only the first time (to download the AI auditor). After that, checks work offline.</p></details>
172
  <details><summary>I'm setting this up for a company</summary><p>Before running, edit <code>baseline.conf</code> (approved admins, approved AI tools, rules) and <code>organisational.conf</code> (8 company questions). They are in <span class="path">~/Library/Application Support/Devseis Endpoint Auditor</span> on Mac, <span class="path">%APPDATA%\Devseis Endpoint Auditor</span> on Windows and <span class="path">~/.config/Devseis Endpoint Auditor</span> on Linux.</p></details>
173
- <details><summary>What about my phone?</summary><p>Open <a href="https://huggingface.co/spaces/Devseis/endpoint-auditor">this page</a> on your iPhone or Android phone — the phone check runs in the browser, nothing to install.</p></details>
174
  <p class="muted">Not legal advice or a certification. Open source by Devseis: code Apache-2.0, data CC BY 4.0. <a href="https://huggingface.co/spaces/Devseis/endpoint-auditor/tree/main">Source</a> · <a href="https://huggingface.co/datasets/Devseis/endpoint-auditor-releases">All releases</a></p>
175
  </main>
176
- <footer class="foot">Devseis · <a href="index.html">Phone check</a> · <a href="about.html">About</a></footer>
177
 
178
  <script>
179
  // Pick the visitor's system, and fill in links, sizes and checksums from the releases repository.
 
69
  <body>
70
  <header class="top">
71
  <div class="brand"><img src="app/build/icon.png" alt=""><span>Devseis <b>Endpoint Auditor</b></span></div>
72
+ <nav><a href="index.html">Assistant</a><a href="phone.html">Phone check</a><a href="about.html">About</a></nav>
73
  </header>
74
 
75
  <main>
 
170
  <details><summary>What is in the report?</summary><p>A score, the most important gaps, and for each problem: what was found, why it matters, how risky it is, and simple steps to fix it — mapped to ISO 27001, GDPR and the EU AI Act.</p></details>
171
  <details><summary>Do I need internet?</summary><p>Only the first time (to download the AI auditor). After that, checks work offline.</p></details>
172
  <details><summary>I'm setting this up for a company</summary><p>Before running, edit <code>baseline.conf</code> (approved admins, approved AI tools, rules) and <code>organisational.conf</code> (8 company questions). They are in <span class="path">~/Library/Application Support/Devseis Endpoint Auditor</span> on Mac, <span class="path">%APPDATA%\Devseis Endpoint Auditor</span> on Windows and <span class="path">~/.config/Devseis Endpoint Auditor</span> on Linux.</p></details>
173
+ <details><summary>What about my phone?</summary><p>Open <a href="phone.html">the phone check</a> on your iPhone or Android phone — the phone check runs in the browser, nothing to install.</p></details>
174
  <p class="muted">Not legal advice or a certification. Open source by Devseis: code Apache-2.0, data CC BY 4.0. <a href="https://huggingface.co/spaces/Devseis/endpoint-auditor/tree/main">Source</a> · <a href="https://huggingface.co/datasets/Devseis/endpoint-auditor-releases">All releases</a></p>
175
  </main>
176
+ <footer class="foot">Devseis · <a href="index.html">Assistant</a> · <a href="phone.html">Phone check</a> · <a href="about.html">About</a></footer>
177
 
178
  <script>
179
  // Pick the visitor's system, and fill in links, sizes and checksums from the releases repository.
index.html CHANGED
@@ -2,91 +2,81 @@
2
  <html lang="en">
3
  <head>
4
  <meta charset="utf-8">
5
- <meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
6
- <meta name="description" content="Free, private security and compliance check for your phone (ISO 27001, GDPR, EU AI Act). Runs in your browser; nothing about your phone is uploaded.">
7
  <meta name="theme-color" content="#ED2939">
8
- <title>Devseis Phone Check</title>
9
- <link rel="stylesheet" href="mobile/styles.css">
 
 
10
  </head>
11
  <body>
12
- <header class="top">
13
- <div class="brand"><span class="mark">D</span><span>Devseis <b>Phone Check</b></span></div>
14
- <div id="model-chip" class="chip hidden"></div>
 
 
 
 
 
 
 
 
 
 
 
15
  </header>
 
16
 
17
- <main>
18
- <!-- 1. Start -->
19
- <section id="s-start" class="screen">
20
- <h1 id="hero-title">Is this phone ready for work?</h1>
21
- <p class="lead" id="hero-lead">A free check against <b>ISO 27001</b>, <b>GDPR</b> and the <b>EU AI Act</b>, written by the Devseis AI auditor <b>inside your browser</b>.</p>
22
- <div id="device-card" class="card device"><span class="muted">Detecting this device…</span></div>
23
- <div id="desktop-note" class="card note hidden">
24
- <p><a class="dl-mac" id="dl-desktop" href="download.html">Download the desktop app for Mac</a></p>
25
- <p class="muted" id="desktop-other"></p>
26
- <p class="muted">The phone check runs on phones: open this page on your iPhone or Android. To try it here as a demo, choose a phone type:</p>
27
- <div class="row"><button class="secondary" data-demo="ios">iPhone</button><button class="secondary" data-demo="android">Android</button></div>
28
  </div>
29
- <ul class="facts" id="facts">
30
- <li><b>Private:</b> your answers and the report stay on this phone. Only the AI model and vulnerability data are downloaded.</li>
31
- <li><b>About 5 minutes:</b> the browser detects what it can; you answer the rest, with help on where to look.</li>
32
- <li><b>Honest:</b> every result says whether it was <i>detected</i> or <i>self-reported</i>.</li>
33
- </ul>
34
- <button id="start" class="primary" disabled>Start the check</button>
35
- <p class="muted center"><a href="about.html">About the project</a></p>
36
- </section>
37
-
38
- <!-- 2. Approval -->
39
- <section id="s-consent" class="screen hidden">
40
- <h1>Before we start</h1>
41
- <div class="card">
42
- <p><b>What will be checked</b> (<span id="check-count"></span> checks):</p>
43
- <ul id="plan" class="plan"></ul>
44
- </div>
45
- <label class="card toggle">
46
- <input type="checkbox" id="use-ai">
47
- <span><b>Use the AI auditor</b><br><span id="ai-note" class="muted"></span></span>
48
- </label>
49
- <p class="muted">Nothing on the phone is changed. You can stop at any time.</p>
50
- <div class="row"><button class="secondary" id="consent-back">Back</button><button class="primary" id="consent-ok">I agree, continue</button></div>
51
- </section>
52
 
53
- <!-- 3. Questions -->
54
- <section id="s-question" class="screen hidden">
55
- <div class="progress"><div id="q-bar" class="bar"></div></div>
56
- <p class="muted" id="q-count"></p>
57
- <h2 id="q-check" class="check-name"></h2>
58
- <h1 id="q-text"></h1>
59
- <div class="where"><span class="where-label">Where to look</span><span id="q-where"></span></div>
60
- <div id="q-answers" class="answers"></div>
61
- <div class="row"><button class="secondary" id="q-back">Back</button></div>
62
- </section>
 
 
 
63
 
64
- <!-- 4. Writing the report -->
65
- <section id="s-work" class="screen hidden">
66
- <h1 id="w-title">Writing your report</h1>
67
- <p id="w-step" class="muted"></p>
68
- <div class="progress big"><div id="w-bar" class="bar"></div></div>
69
- <p id="w-current" class="muted"></p>
70
- </section>
 
 
 
 
71
 
72
- <!-- 5. Result -->
73
- <section id="s-done" class="screen hidden">
74
- <div id="result"></div>
75
- <div class="actions">
76
- <button class="primary" id="share-pdf">Save or share PDF</button>
77
- <button class="secondary" id="view-report">View full report</button>
78
- <button class="secondary" id="save-json">Download evidence (JSON)</button>
79
- </div>
80
- <p id="pdf-note" class="muted"></p>
81
- <button class="link" id="again">Check again</button>
82
- </section>
83
- </main>
84
 
85
- <footer class="foot">
86
- Open source by Devseis · <a href="about.html">about</a> ·
87
- <a href="https://huggingface.co/spaces/Devseis/endpoint-auditor/tree/main">code (Apache-2.0)</a> ·
88
- <a href="https://huggingface.co/datasets/Devseis/endpoint-auditor-synthetic">data (CC BY 4.0)</a> · Not legal advice or a certification.
89
- </footer>
90
- <script type="module" src="mobile/app.js"></script>
91
  </body>
92
  </html>
 
2
  <html lang="en">
3
  <head>
4
  <meta charset="utf-8">
5
+ <meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover, interactive-widget=resizes-content">
6
+ <meta name="description" content="Devseis security assistant: ask about securing your computer or phone, ISO 27001, GDPR and the EU AI Act, and run a free audit. The AI runs on your device; nothing is uploaded.">
7
  <meta name="theme-color" content="#ED2939">
8
+ <title>Devseis Security Assistant</title>
9
+ <link rel="icon" href="app/build/icon.png">
10
+ <link rel="apple-touch-icon" href="app/build/icon.png">
11
+ <link rel="stylesheet" href="chat/chat.css">
12
  </head>
13
  <body>
14
+ <header>
15
+ <div class="header-left">
16
+ <button type="button" id="sidebarToggle" class="icon-btn" aria-label="Conversations" title="Conversations">
17
+ <svg viewBox="0 0 24 24" width="20" height="20" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round"><path d="M4 6h16M4 12h16M4 18h16"/></svg>
18
+ </button>
19
+ <a class="brand" href="index.html"><img src="app/build/icon.png" alt=""><span>Devseis <b>Security Assistant</b></span></a>
20
+ </div>
21
+ <div class="header-actions">
22
+ <button type="button" id="newChat" class="icon-btn" aria-label="New chat" title="New chat">
23
+ <svg viewBox="0 0 24 24" width="19" height="19" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round"><path d="M12 5v14M5 12h14"/></svg>
24
+ </button>
25
+ <button type="button" id="themeToggle" class="icon-btn" aria-label="Light or dark" title="Light or dark">◐</button>
26
+ <button type="button" id="auditBtn" class="audit-btn">🛡 Run audit</button>
27
+ </div>
28
  </header>
29
+ <div class="status-bar hidden" id="statusBar"><div class="status-text" id="statusText"></div><div class="status-progress"><div id="statusFill"></div></div></div>
30
 
31
+ <main id="chatMain">
32
+ <div class="empty-state" id="emptyState">
33
+ <div class="hero-icon"><img src="app/build/icon.png" alt=""></div>
34
+ <h1>Hi, I'm your Devseis security assistant.</h1>
35
+ <p class="hero-sub">Ask me how to secure your computer or phone, or about ISO 27001, GDPR and the EU AI Act. I can also <b>audit this device</b> for free. I run on your device: nothing you type is uploaded.</p>
36
+ <div class="chips">
37
+ <button type="button" class="chip" data-prompt="How do I turn on disk encryption on my computer?">🔒 Turn on disk encryption</button>
38
+ <button type="button" class="chip" data-prompt="Is ChatGPT OK to use at work under GDPR?">🤖 Is ChatGPT OK at work?</button>
39
+ <button type="button" class="chip" data-prompt="What does ISO 27001 A.8.7 require?">📘 Explain ISO 27001 A.8.7</button>
40
+ <button type="button" class="chip" data-action="audit">🛡 Audit this device</button>
 
41
  </div>
42
+ </div>
43
+ <div id="thread" aria-live="polite"></div>
44
+ </main>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45
 
46
+ <form id="composer" autocomplete="off">
47
+ <button type="button" id="attachBtn" class="icon-btn attach" title="Add an audit report (.json from the Devseis app)" aria-label="Add an audit report">
48
+ <svg viewBox="0 0 24 24" width="19" height="19" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M21.44 11.05l-9.19 9.19a5.5 5.5 0 01-7.78-7.78l9.2-9.19a3.5 3.5 0 014.95 4.95l-9.2 9.19a1.5 1.5 0 01-2.12-2.12l8.49-8.48"/></svg>
49
+ </button>
50
+ <input type="file" id="fileInput" accept=".json,application/json" hidden>
51
+ <div class="input-pill">
52
+ <textarea id="input" rows="1" placeholder="Ask about security, or add your audit report…" aria-label="Message"></textarea>
53
+ </div>
54
+ <button type="submit" id="sendBtn" class="send-btn" aria-label="Send" disabled>
55
+ <svg viewBox="0 0 24 24" width="18" height="18" fill="currentColor"><path d="M3.4 20.4l17.45-7.48a1 1 0 000-1.84L3.4 3.6a.99.99 0 00-1.39.91L2 9.12c0 .5.37.93.87.99L17 12 2.87 13.88c-.5.07-.87.5-.87 1l.01 4.61c0 .71.73 1.2 1.39.91z"/></svg>
56
+ </button>
57
+ </form>
58
+ <p class="disclaimer">Runs on your device · Answers come from the Devseis knowledge base and can be wrong · Not legal advice</p>
59
 
60
+ <div class="overlay hidden" id="overlay"></div>
61
+ <aside class="sidebar" id="sidebar" aria-label="Conversations">
62
+ <div class="sidebar-head"><h2>Conversations</h2><button type="button" class="icon-btn" id="sidebarClose" aria-label="Close">✕</button></div>
63
+ <div class="sidebar-list" id="sidebarList"></div>
64
+ <div class="sidebar-foot">
65
+ <div class="row"><span>AI model</span><span class="muted" id="modelName">loads on first question</span></div>
66
+ <div class="row"><a href="phone.html">📱 Phone check</a><a href="download.html">💻 Desktop app</a></div>
67
+ <div class="row"><a href="about.html">About</a><button type="button" class="small-btn" id="clearAll">Clear history</button></div>
68
+ <p class="muted tiny">Conversations are stored only in this browser. Open source by Devseis (Apache-2.0).</p>
69
+ </div>
70
+ </aside>
71
 
72
+ <div class="modal hidden" id="auditModal" role="dialog" aria-modal="true" aria-labelledby="auditTitle">
73
+ <div class="modal-card">
74
+ <h2 id="auditTitle">Audit this device</h2>
75
+ <div id="auditOptions"></div>
76
+ <button type="button" class="link-btn" id="auditClose">Not now</button>
77
+ </div>
78
+ </div>
 
 
 
 
 
79
 
80
+ <script type="module" src="chat/chat.js"></script>
 
 
 
 
 
81
  </body>
82
  </html>
library/build_knowledge.py ADDED
@@ -0,0 +1,184 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Build library/knowledge.json: the passages the chat assistant answers from (and cites).
3
+
4
+ Generated from the catalog and library so the chat never drifts from the audit:
5
+ fix how to fix / why it matters, per check and OS (check_guidance.json + catalog rule + references)
6
+ control ISO 27001 / GDPR / EU AI Act control summaries (controls.json)
7
+ tool AI tool profiles (ai_tools.json) and EU AI Act risk categories
8
+ phone where a setting is on iPhone / Android (catalog/mobile_questions.json)
9
+ status what each audit status means
10
+ faq the auditor itself: what it checks, privacy, install, phone check
11
+ Usage: python3 library/build_knowledge.py
12
+ """
13
+ import json
14
+ import os
15
+ import re
16
+ import sys
17
+
18
+ ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
19
+ OS_LABEL = {"windows": "Windows", "linux": "Linux", "macos": "macOS", "ios": "iPhone", "android": "Android"}
20
+
21
+
22
+ def load(*parts):
23
+ with open(os.path.join(ROOT, *parts), encoding="utf-8") as handle:
24
+ return json.load(handle)
25
+
26
+
27
+ def read_conf(path):
28
+ conf = {}
29
+ for line in open(os.path.join(ROOT, path), encoding="utf-8"):
30
+ if "=" in line and not line.lstrip().startswith("#"):
31
+ key, value = line.split("=", 1)
32
+ conf[key.strip()] = value.strip().strip('"')
33
+ return conf
34
+
35
+
36
+ def fill(text, baseline):
37
+ """Same placeholder rules as the report builder (training/build_dataset.py), including derived values."""
38
+ sys.path.insert(0, os.path.join(ROOT, "training"))
39
+ from build_dataset import fill as report_fill
40
+ return report_fill(text, baseline)
41
+
42
+
43
+ FAQ = [
44
+ ("What the Devseis Endpoint Auditor is",
45
+ "The Devseis Endpoint Auditor is a free, open-source, read-only check of a computer or phone against ISO 27001, GDPR "
46
+ "technical controls and the EU AI Act. It reads security settings, compares them with a baseline, matches installed "
47
+ "software against known vulnerabilities, and an AI auditor that runs on the device writes a report with findings, "
48
+ "risk levels and fix steps. It is not a certification and not legal advice."),
49
+ ("Does the audit change anything?",
50
+ "No. The audit only reads settings such as password rules, firewall, disk encryption and installed software. It never "
51
+ "changes them. Administrator rights are requested because some settings can only be read by an administrator."),
52
+ ("Privacy: where do audit results go?",
53
+ "Nowhere. Evidence and reports stay on the device. The app downloads only two things: the AI auditor model and a signed "
54
+ "list of known vulnerabilities. Nothing about the device is uploaded. Chat history is stored only in this browser."),
55
+ ("How to run an audit of a computer",
56
+ "A web page cannot read a computer's settings, so computers are audited with the free Devseis desktop app for Mac, "
57
+ "Windows (beta) and Linux. Download it from the Download page, install it, click Run audit and enter the computer "
58
+ "password once. The first run downloads the AI auditor; the audit then takes about 5 to 15 minutes and ends with a PDF report."),
59
+ ("How to install on a Mac",
60
+ "Open the downloaded .dmg file and drag Devseis Endpoint Auditor to Applications. The first time, macOS says it could not "
61
+ "verify the app because it is not yet notarised: click Done, then open System Settings, Privacy & Security, and click "
62
+ "Open Anyway. After that it opens normally. It needs macOS 13 or newer."),
63
+ ("How to install on Windows",
64
+ "Run the downloaded setup .exe. Windows SmartScreen shows 'Windows protected your PC' because the app is not yet "
65
+ "code-signed: click More info, then Run anyway. Click Next, Install and Finish. When the audit starts, click Yes in the "
66
+ "User Account Control prompt. The Windows version is a beta."),
67
+ ("How to install on Linux",
68
+ "On Ubuntu, Debian, Linux Mint or Pop!_OS, double-click the .deb file and click Install. On other distributions, make "
69
+ "the AppImage executable (Properties, Permissions, Allow executing file as program) and double-click it. The audit asks "
70
+ "for the password with the normal desktop prompt."),
71
+ ("How the phone check works",
72
+ "Phones are checked in the browser: open the Devseis page on the iPhone or Android phone and start the phone check. The "
73
+ "browser detects what it can (system version, whether a screen lock is set up) and asks the rest with help on where to "
74
+ "look in Settings. Every result is labelled detected or self-reported. The report can be saved or shared as a PDF."),
75
+ ("What self-reported and detected mean",
76
+ "Detected means the browser or the app read the value itself. Self-reported means the user answered a question; the "
77
+ "value was not verified on the device. Phone checks are mostly self-reported because browsers cannot read phone settings."),
78
+ ("What is in the audit report",
79
+ "The report has a compliance score, the most important gaps, and for each problem: what was found, why it matters, the "
80
+ "risk level, step-by-step fixes for that operating system, how to verify the fix, and the related ISO 27001, GDPR and EU "
81
+ "AI Act references. It also lists AI tools found and their EU AI Act risk class."),
82
+ ("Does the auditor need internet?",
83
+ "Only the first time, to download the AI auditor and the vulnerability list. After that audits and chat work offline. "
84
+ "Pending security updates can only be checked online because they ask the vendor."),
85
+ ("Settings for organisations",
86
+ "Before running audits, an organisation can edit baseline.conf (approved administrator accounts, approved AI tools, "
87
+ "limits such as screen lock within 5 minutes) and organisational.conf (answers to 8 organisational questions such as "
88
+ "record of processing, DPIA and breach procedure). The desktop app keeps them in its settings folder."),
89
+ ("The Devseis AI auditor model",
90
+ "The assistant and the report writer are a small language model fine-tuned by Devseis from Qwen2.5 and run on the "
91
+ "device with WebLLM. Audit verdicts are decided by rules in code, not by the model; the model explains, prioritises and "
92
+ "writes, and every report answer is checked against the evidence before it is used."),
93
+ ]
94
+
95
+
96
+ def main():
97
+ catalog = load("catalog", "checks.json")
98
+ guidance = load("library", "check_guidance.json")
99
+ controls = load("library", "controls.json")
100
+ tools = load("library", "ai_tools.json")
101
+ phone = load("catalog", "mobile_questions.json")
102
+ baseline = read_conf("baseline.conf")
103
+ refs = catalog["references"]
104
+ ctl = {(c["framework"], c["id"]): c for c in controls["controls"]}
105
+ chunks = []
106
+
107
+ def add(cid, kind, title, text, os_family=None, check=None):
108
+ chunks.append({"id": cid, "kind": kind, "title": title, "text": re.sub(r"\s+", " ", text).strip(),
109
+ **({"os": os_family} if os_family else {}), **({"check": check} if check else {})})
110
+
111
+ def ref_text(check):
112
+ out = []
113
+ for key, label in (("iso27001", "ISO 27001:2022"), ("gdpr", "GDPR"), ("ai_act", "EU AI Act")):
114
+ for r in check["controls"][key]:
115
+ title = (ctl.get((label, r)) or {}).get("title") or refs[key].get(r, "")
116
+ out.append(f"{label} {r}" + (f" ({title})" if title else ""))
117
+ return "; ".join(out)
118
+
119
+ for check in catalog["checks"]:
120
+ if check["section"] == "org":
121
+ g = guidance["checks"].get(guidance["aliases"].get(check["id"], check["id"]))
122
+ add(f"fix:{check['id']}", "fix", check["name"],
123
+ f"{check['name']} is an organisational control, answered once by the organisation, not read from a device. "
124
+ f"{g['why'] if g else ''} Rule: {check['rule']} Related: {ref_text(check)}.", check=check["id"])
125
+ continue
126
+ g = guidance["checks"].get(guidance["aliases"].get(check["id"], check["id"]))
127
+ if not g:
128
+ continue
129
+ for os_family in check["os"]:
130
+ name = (check.get("names") or {}).get(os_family, check["name"])
131
+ steps = g["fix"].get("all", []) + g["fix"].get(os_family, [])
132
+ text = (f"{name} on {OS_LABEL[os_family]}. Why it matters: {g['why']} Risk if missing: {g['risk']['level']} - "
133
+ f"{g['risk']['reason']} Requirement: {fill(check['rule'], baseline)} How to fix on {OS_LABEL[os_family]}: "
134
+ + " ".join(f"({i}) {fill(s, baseline)}" for i, s in enumerate(steps, 1))
135
+ + f" How to verify: {fill(g['verify'], baseline)} Related controls: {ref_text(check)}.")
136
+ add(f"fix:{check['id']}:{os_family}", "fix", f"{name} on {OS_LABEL[os_family]}", text, os_family, check["id"])
137
+
138
+ for c in controls["controls"]:
139
+ add(f"control:{c['framework']}:{c['id']}", "control", f"{c['framework']} {c['id']} {c['title']}",
140
+ f"{c['framework']} {c['id']} ({c['title']}): {c['summary']} An auditor looks for: {'; '.join(c.get('auditor_looks_for', []))}. "
141
+ f"On endpoints: {c.get('endpoint_relevance', '')}")
142
+
143
+ for name, t in tools["tools"].items():
144
+ add(f"tool:{name}", "tool", f"{name} (AI tool)",
145
+ f"{name}: {t['kind']}, provider {t['provider']}, data location: {t['data_location']}. Default EU AI Act risk class "
146
+ f"for general office use: {t['default_risk']}. GDPR: {t['gdpr_notes']} Whether it is allowed at work depends on the "
147
+ f"organisation's approved AI tools list and on the purpose: some purposes make any AI tool high-risk or prohibited.")
148
+ rules = tools["purpose_rules"]
149
+ for key, title in (("prohibited", "EU AI Act prohibited AI practices"), ("high_risk_examples", "EU AI Act high-risk AI uses"),
150
+ ("limited_risk", "EU AI Act limited-risk AI (transparency duties)"), ("minimal_risk", "EU AI Act minimal-risk AI")):
151
+ items = rules[key] if isinstance(rules[key], list) else [rules[key]]
152
+ add(f"aiact:{key}", "control", title, f"{title}: " + "; ".join(items) + ".")
153
+
154
+ for check_id, per_os in phone["checks"].items():
155
+ for os_family, spec in per_os.items():
156
+ name = next(((c.get("names") or {}).get(os_family, c["name"]) for c in catalog["checks"] if c["id"] == check_id), check_id)
157
+ add(f"phone:{check_id}:{os_family}", "phone", f"Where to find {name} on {OS_LABEL[os_family]}",
158
+ f"{name} on {OS_LABEL[os_family]}: {spec['question']} Where to look: {spec['where']}", os_family, check_id)
159
+
160
+ notes = guidance["status_notes"]
161
+ add("status:all", "status", "What the audit statuses mean",
162
+ "Compliant: the setting meets the baseline. Non-Compliant: it does not meet the baseline. "
163
+ f"Pending: {notes['Pending']} NotApplicable: {notes['NotApplicable']} Error: {notes['Error']}")
164
+
165
+ counts = {o: sum(o in c["os"] for c in catalog["checks"]) for o in OS_LABEL}
166
+ add("faq:checks", "faq", "How many checks the audit runs",
167
+ "The audit catalog has " + str(len(catalog["checks"])) + " checks: " + ", ".join(f"{OS_LABEL[o]} {n}" for o, n in counts.items())
168
+ + ". They cover passwords and lockout, malware protection, firewall, screen lock, USB storage, administrator accounts, "
169
+ "disk encryption, updates and known vulnerabilities, GDPR technical controls (backups, logging, remote access, "
170
+ "telemetry and more), EU AI Act controls (AI tools found, shadow AI, risk classification) and 8 organisational controls.")
171
+ for i, (title, text) in enumerate(FAQ):
172
+ add(f"faq:{i}", "faq", title, text)
173
+
174
+ out = {"version": catalog["catalog_version"], "chunks": chunks}
175
+ with open(os.path.join(ROOT, "library", "knowledge.json"), "w", encoding="utf-8") as handle:
176
+ json.dump(out, handle, ensure_ascii=False, indent=0)
177
+ kinds = {}
178
+ for c in chunks:
179
+ kinds[c["kind"]] = kinds.get(c["kind"], 0) + 1
180
+ print(f"knowledge.json: {len(chunks)} passages {kinds}")
181
+
182
+
183
+ if __name__ == "__main__":
184
+ main()
library/knowledge.json ADDED
The diff for this file is too large to render. See raw diff
 
mobile/app.js CHANGED
@@ -268,6 +268,10 @@ function showResult() {
268
  <p class="muted">Written ${sources.model ? `by the AI auditor (${sources.model}) and ` : ''}by the built-in writer (${sources.template}).</p>
269
  </div>`;
270
  $('pdf-note').textContent = 'The report stays on this phone until you save or share it.';
 
 
 
 
271
  show('done');
272
  }
273
 
 
268
  <p class="muted">Written ${sources.model ? `by the AI auditor (${sources.model}) and ` : ''}by the built-in writer (${sources.template}).</p>
269
  </div>`;
270
  $('pdf-note').textContent = 'The report stays on this phone until you save or share it.';
271
+ try { localStorage.setItem('devseis.lastAudit', state.report.evidenceText); } catch { /* private mode */ }
272
+ if (!document.getElementById('ask-assistant')) {
273
+ document.querySelector('#s-done .actions').insertAdjacentHTML('afterbegin', '<a class="dl-mac" id="ask-assistant" href="index.html#audit">💬 Ask the assistant about this report</a>');
274
+ }
275
  show('done');
276
  }
277
 
mobile/test/e2e.cjs CHANGED
@@ -28,7 +28,7 @@ app.whenReady().then(async () => {
28
  const until = async (code, ms = 60000) => { const t = Date.now(); while (!(await js(code))) { if (Date.now() - t > ms) throw new Error(`timeout: ${code}`); await new Promise(r => setTimeout(r, 200)); } };
29
  try {
30
  await win.webContents.session.clearCache();
31
- await win.loadURL(`${process.env.BASE_URL || 'http://127.0.0.1:8765/'}index.html?test`);
32
  await until(`!document.getElementById('start').disabled`);
33
  const device = await js(`document.getElementById('device-card').innerText`);
34
  await js(`document.getElementById('start').click()`);
@@ -48,6 +48,7 @@ app.whenReady().then(async () => {
48
  const evidence = await js(`window.__devseisState.report.evidenceText`);
49
  const result = await js(`document.getElementById('result').innerText`);
50
  const pdfSize = await js(`window.__devseisMakePdf().then(b => b.size)`);
 
51
  const html = await js(`window.__devseisState.report.html`);
52
  fs.writeFileSync(out, evidence);
53
  fs.writeFileSync(out.replace(/\.json$/, '.html'), html);
 
28
  const until = async (code, ms = 60000) => { const t = Date.now(); while (!(await js(code))) { if (Date.now() - t > ms) throw new Error(`timeout: ${code}`); await new Promise(r => setTimeout(r, 200)); } };
29
  try {
30
  await win.webContents.session.clearCache();
31
+ await win.loadURL(`${process.env.BASE_URL || 'http://127.0.0.1:8765/'}phone.html?test`);
32
  await until(`!document.getElementById('start').disabled`);
33
  const device = await js(`document.getElementById('device-card').innerText`);
34
  await js(`document.getElementById('start').click()`);
 
48
  const evidence = await js(`window.__devseisState.report.evidenceText`);
49
  const result = await js(`document.getElementById('result').innerText`);
50
  const pdfSize = await js(`window.__devseisMakePdf().then(b => b.size)`);
51
+ console.log('handoff:', await js(`(document.getElementById('ask-assistant')||{}).href + ' stored=' + !!localStorage.getItem('devseis.lastAudit')`));
52
  const html = await js(`window.__devseisState.report.html`);
53
  fs.writeFileSync(out, evidence);
54
  fs.writeFileSync(out.replace(/\.json$/, '.html'), html);
phone.html ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8">
5
+ <meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
6
+ <meta name="description" content="Free, private security and compliance check for your phone (ISO 27001, GDPR, EU AI Act). Runs in your browser; nothing about your phone is uploaded.">
7
+ <meta name="theme-color" content="#ED2939">
8
+ <title>Devseis Phone Check</title>
9
+ <link rel="stylesheet" href="mobile/styles.css">
10
+ </head>
11
+ <body>
12
+ <header class="top">
13
+ <div class="brand"><span class="mark">D</span><span>Devseis <b>Phone Check</b></span></div>
14
+ <div style="display:flex;gap:8px;align-items:center"><div id="model-chip" class="chip hidden"></div><a class="chip" href="index.html">← Assistant</a></div>
15
+ </header>
16
+
17
+ <main>
18
+ <!-- 1. Start -->
19
+ <section id="s-start" class="screen">
20
+ <h1 id="hero-title">Is this phone ready for work?</h1>
21
+ <p class="lead" id="hero-lead">A free check against <b>ISO 27001</b>, <b>GDPR</b> and the <b>EU AI Act</b>, written by the Devseis AI auditor <b>inside your browser</b>.</p>
22
+ <div id="device-card" class="card device"><span class="muted">Detecting this device…</span></div>
23
+ <div id="desktop-note" class="card note hidden">
24
+ <p><a class="dl-mac" id="dl-desktop" href="download.html">Download the desktop app for Mac</a></p>
25
+ <p class="muted" id="desktop-other"></p>
26
+ <p class="muted">The phone check runs on phones: open this page on your iPhone or Android. To try it here as a demo, choose a phone type:</p>
27
+ <div class="row"><button class="secondary" data-demo="ios">iPhone</button><button class="secondary" data-demo="android">Android</button></div>
28
+ </div>
29
+ <ul class="facts" id="facts">
30
+ <li><b>Private:</b> your answers and the report stay on this phone. Only the AI model and vulnerability data are downloaded.</li>
31
+ <li><b>About 5 minutes:</b> the browser detects what it can; you answer the rest, with help on where to look.</li>
32
+ <li><b>Honest:</b> every result says whether it was <i>detected</i> or <i>self-reported</i>.</li>
33
+ </ul>
34
+ <button id="start" class="primary" disabled>Start the check</button>
35
+ <p class="muted center"><a href="about.html">About the project</a></p>
36
+ </section>
37
+
38
+ <!-- 2. Approval -->
39
+ <section id="s-consent" class="screen hidden">
40
+ <h1>Before we start</h1>
41
+ <div class="card">
42
+ <p><b>What will be checked</b> (<span id="check-count"></span> checks):</p>
43
+ <ul id="plan" class="plan"></ul>
44
+ </div>
45
+ <label class="card toggle">
46
+ <input type="checkbox" id="use-ai">
47
+ <span><b>Use the AI auditor</b><br><span id="ai-note" class="muted"></span></span>
48
+ </label>
49
+ <p class="muted">Nothing on the phone is changed. You can stop at any time.</p>
50
+ <div class="row"><button class="secondary" id="consent-back">Back</button><button class="primary" id="consent-ok">I agree, continue</button></div>
51
+ </section>
52
+
53
+ <!-- 3. Questions -->
54
+ <section id="s-question" class="screen hidden">
55
+ <div class="progress"><div id="q-bar" class="bar"></div></div>
56
+ <p class="muted" id="q-count"></p>
57
+ <h2 id="q-check" class="check-name"></h2>
58
+ <h1 id="q-text"></h1>
59
+ <div class="where"><span class="where-label">Where to look</span><span id="q-where"></span></div>
60
+ <div id="q-answers" class="answers"></div>
61
+ <div class="row"><button class="secondary" id="q-back">Back</button></div>
62
+ </section>
63
+
64
+ <!-- 4. Writing the report -->
65
+ <section id="s-work" class="screen hidden">
66
+ <h1 id="w-title">Writing your report</h1>
67
+ <p id="w-step" class="muted"></p>
68
+ <div class="progress big"><div id="w-bar" class="bar"></div></div>
69
+ <p id="w-current" class="muted"></p>
70
+ </section>
71
+
72
+ <!-- 5. Result -->
73
+ <section id="s-done" class="screen hidden">
74
+ <div id="result"></div>
75
+ <div class="actions">
76
+ <button class="primary" id="share-pdf">Save or share PDF</button>
77
+ <button class="secondary" id="view-report">View full report</button>
78
+ <button class="secondary" id="save-json">Download evidence (JSON)</button>
79
+ </div>
80
+ <p id="pdf-note" class="muted"></p>
81
+ <button class="link" id="again">Check again</button>
82
+ </section>
83
+ </main>
84
+
85
+ <footer class="foot">
86
+ Open source by Devseis · <a href="about.html">about</a> ·
87
+ <a href="https://huggingface.co/spaces/Devseis/endpoint-auditor/tree/main">code (Apache-2.0)</a> ·
88
+ <a href="https://huggingface.co/datasets/Devseis/endpoint-auditor-synthetic">data (CC BY 4.0)</a> · Not legal advice or a certification.
89
+ </footer>
90
+ <script type="module" src="mobile/app.js"></script>
91
+ </body>
92
+ </html>
training/build_chat.py ADDED
@@ -0,0 +1,297 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Chat training examples for model v0.4 (questions about endpoint security, ISO 27001, GDPR, EU AI Act, the auditor,
3
+ and the user's own audit), grounded in library/knowledge.json and cited as [n].
4
+
5
+ Answers are composed from the same library facts the passages are built from, so they are grounded by construction;
6
+ check_chat() verifies it (citations exist, numbers appear in the prompt, refusals cite nothing). Prompts are built by
7
+ training/chat-contexts.mjs with the app's own search, so training matches runtime.
8
+
9
+ Usage: python3 training/build_chat.py --evidence synthetic/out-v0.3/evidence.jsonl --base training/out-v0.3 --out training/out-v0.4
10
+ """
11
+ import argparse
12
+ import json
13
+ import os
14
+ import random
15
+ import re
16
+ import subprocess
17
+ import sys
18
+
19
+ ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
20
+ sys.path.insert(0, os.path.join(ROOT, "library"))
21
+ import build_knowledge as kb_build # noqa: E402
22
+
23
+ DATASET_VERSION = "0.4.0"
24
+ OS_LABEL = kb_build.OS_LABEL
25
+ NUMBER = re.compile(r"(?<![\w.])\d+(?:\.\d+)?(?!\w)")
26
+
27
+
28
+ def load(*parts):
29
+ with open(os.path.join(ROOT, *parts), encoding="utf-8") as handle:
30
+ return json.load(handle)
31
+
32
+
33
+ def cite(order, cid):
34
+ return f"[{order.index(cid) + 1}]"
35
+
36
+
37
+ def steps_text(steps):
38
+ return "\n".join(f"{i}. {s}" for i, s in enumerate(steps, 1))
39
+
40
+
41
+ OUT_OF_SCOPE = ["What's a good pizza recipe?", "Write me a poem about the sea.", "Who won the football match last night?",
42
+ "Can you help me with my maths homework?", "What's the weather like tomorrow?", "Tell me a joke.",
43
+ "Translate 'good morning' into Spanish.", "Which phone should I buy?", "How do I lose weight?",
44
+ "Recommend a good film for tonight.", "Write a cover letter for a sales job.", "What is the capital of Australia?",
45
+ "How do I make sourdough bread?", "Summarise the plot of Hamlet.", "What stocks should I buy?",
46
+ "Plan a weekend trip to Paris.", "How do I fix a leaking tap?", "Write Python code to sort a list.",
47
+ "What's the best gaming laptop?", "Explain how black holes form.", "Help me write a birthday message.",
48
+ "Who is the president of France?", "How many calories are in a banana?", "Give me a workout plan."]
49
+ UNKNOWN = ["How do I configure VLANs on a Cisco switch?", "What is the PCI DSS requirement 8.3.6?", "How do I set up SSO with Okta?",
50
+ "Is our AWS S3 bucket policy secure?", "How do I configure DMARC for my domain?", "What does SOC 2 Type II require for change management?",
51
+ "How do I harden a Kubernetes cluster?", "Which SIEM should we buy?", "What does HIPAA say about encryption?",
52
+ "How do I configure Azure Conditional Access?", "What does NIS2 require from my company?", "How do I rotate keys in HashiCorp Vault?",
53
+ "Is our WordPress site secure?", "How do I set up a WAF in Cloudflare?", "What are the DORA ICT risk rules?",
54
+ "How do I scan my network with nmap?"]
55
+ OWN_DEVICE = ["Is my laptop secure?", "Is my firewall OK?", "Is my computer compliant with ISO 27001?", "What should I fix first?",
56
+ "Am I GDPR compliant?", "Does my phone have any security problems?", "Is my disk encrypted?", "Are my passwords strong enough?",
57
+ "Do I have any viruses?", "Is my screen lock OK?", "Which of my apps are outdated?", "Is my Mac up to date?",
58
+ "Am I using any unapproved AI tools?", "What is my compliance score?"]
59
+ SCOPE = ("I can only help with device security, ISO 27001, GDPR and the EU AI Act, so I can't answer that. "
60
+ "Ask me, for example, how to turn on disk encryption, whether an AI tool is OK for work, or run an audit of this device.")
61
+ DONT_KNOW = ("I don't have reliable information about that in my knowledge base, so I won't guess. I cover security settings "
62
+ "on Windows, macOS, Linux, iPhone and Android, ISO 27001, GDPR and the EU AI Act.")
63
+
64
+
65
+ def make_jobs(rng, catalog, guidance, controls, tools, phone, evidences):
66
+ """(job for chat-contexts.mjs, answer builder(order) -> text, meta)."""
67
+ jobs = []
68
+ checks = [c for c in catalog["checks"] if c["section"] != "org"]
69
+ G = lambda cid: guidance["checks"][guidance["aliases"].get(cid, cid)]
70
+ refs_of = lambda c: [f"{lab} {r}" for key, lab in (("iso27001", "ISO 27001:2022"), ("gdpr", "GDPR"), ("ai_act", "EU AI Act")) for r in c["controls"][key]]
71
+ baseline = kb_build.read_conf("baseline.conf")
72
+ fill = lambda t: kb_build.fill(t, baseline)
73
+
74
+ for c in checks:
75
+ g = G(c["id"])
76
+ for os_family in c["os"]:
77
+ name = (c.get("names") or {}).get(os_family, c["name"])
78
+ label = OS_LABEL[os_family]
79
+ gold = f"fix:{c['id']}:{os_family}"
80
+ steps = [fill(s) for s in g["fix"].get("all", []) + g["fix"].get(os_family, [])]
81
+ q_fix = rng.choice([f"How do I fix {name.lower()} on {label}?", f"How can I set up {name.lower()} on my {label}?",
82
+ f"My audit says '{name}' is non-compliant on my {label}. What should I do?",
83
+ f"What steps do I take for {name.lower()} on {label}?"])
84
+ jobs.append(({"question": q_fix, "os": os_family, "gold": [gold]},
85
+ lambda o, g=g, steps=steps, gold=gold, label=label: f"{g['why']} To fix it on {label} {cite(o, gold)}:\n{steps_text(steps)}\n\nTo check it worked: {fill(g['verify'])}",
86
+ ("fix", os_family)))
87
+ if rng.random() < 0.45:
88
+ q = rng.choice([f"Why does {name.lower()} matter?", f"What's the risk if {name.lower()} isn't set up on my {label}?",
89
+ f"Is {name.lower()} really important on {label}?"])
90
+ jobs.append(({"question": q, "os": os_family, "gold": [gold]},
91
+ lambda o, g=g, gold=gold: f"{g['why']} If it is missing, the risk is {g['risk']['level'].lower()}: {g['risk']['reason']} {cite(o, gold)}",
92
+ ("why", os_family)))
93
+ if rng.random() < 0.25:
94
+ rl = refs_of(c)
95
+ if rl:
96
+ q = rng.choice([f"Which ISO 27001, GDPR or EU AI Act rules cover {name.lower()}?", f"What compliance controls relate to {name.lower()}?"])
97
+ jobs.append(({"question": q, "os": os_family, "gold": [gold]},
98
+ lambda o, rl=rl, name=name, gold=gold: f"{name} relates to {', '.join(rl)} {cite(o, gold)}. "
99
+ "Meeting the setting is evidence for these controls; it is not a certification on its own.",
100
+ ("refs", os_family)))
101
+ if rng.random() < 0.2:
102
+ q = f"How do I check that {name.lower()} is set correctly on {label}?"
103
+ jobs.append(({"question": q, "os": os_family, "gold": [gold]},
104
+ lambda o, g=g, gold=gold: f"{fill(g['verify'])} {cite(o, gold)}", ("verify", os_family)))
105
+
106
+ for check_id, per_os in phone["checks"].items():
107
+ for os_family, spec in per_os.items():
108
+ name = next(((c.get("names") or {}).get(os_family, c["name"]) for c in catalog["checks"] if c["id"] == check_id), check_id)
109
+ gold = f"phone:{check_id}:{os_family}"
110
+ for q in rng.sample([f"Where do I find {name.lower()} on my {OS_LABEL[os_family]}?", f"How do I see {name.lower()} on {OS_LABEL[os_family]}?",
111
+ f"Where is the {name.lower()} setting on my {OS_LABEL[os_family]}?"], 2):
112
+ jobs.append(({"question": q, "os": os_family, "gold": [gold]},
113
+ lambda o, spec=spec, gold=gold: f"Go to {spec['where']} {cite(o, gold)}", ("phone", os_family)))
114
+
115
+ for c in controls["controls"]:
116
+ gold = f"control:{c['framework']}:{c['id']}"
117
+ name = f"{c['framework'].replace(':2022', '')} {c['id']}"
118
+ for q in rng.sample([f"What does {name} require?", f"Explain {name} ({c['title']}) in simple terms.",
119
+ f"What does an auditor look for under {name}?"], 2):
120
+ jobs.append(({"question": q, "os": None, "gold": [gold]},
121
+ lambda o, c=c, gold=gold: f"{c['id']} ({c['title']}): {c['summary']} {cite(o, gold)}\n\nAn auditor looks for:\n"
122
+ + "\n".join(f"- {x}" for x in c.get("auditor_looks_for", [])) + (f"\n\nOn devices: {c['endpoint_relevance']}" if c.get("endpoint_relevance") else ""),
123
+ ("control", "any")))
124
+
125
+ for name, t in tools["tools"].items():
126
+ gold = f"tool:{name}"
127
+ for q in rng.sample([f"Is {name} OK to use at work?", f"Can I use {name} with customer data under GDPR?",
128
+ f"What is the EU AI Act risk class of {name}?", f"Is {name} safe for my company?"], 3):
129
+ jobs.append(({"question": q, "os": None, "gold": [gold]},
130
+ lambda o, t=t, name=name, gold=gold: f"{name} is a {t['kind'].lower()} from {t['provider']}; data goes to: {t['data_location'].lower()} {cite(o, gold)}. "
131
+ f"For general office use its default EU AI Act risk class is {t['default_risk']}. GDPR: {t['gdpr_notes']}\n\n"
132
+ "Whether it is allowed depends on your organisation's approved AI tools list and on what you use it for: some purposes "
133
+ "(for example screening job applicants) make any AI tool high-risk.", ("tool", "any")))
134
+ rules = tools["purpose_rules"]
135
+ for key, qs in (("prohibited", ["Which AI uses are banned under the EU AI Act?", "What AI practices are prohibited in the EU?", "Is emotion recognition at work allowed?"]),
136
+ ("high_risk_examples", ["When does an AI tool become high-risk under the EU AI Act?", "What are examples of high-risk AI uses?"]),
137
+ ("limited_risk", ["What does limited-risk AI mean under the EU AI Act?"]), ("minimal_risk", ["What is minimal-risk AI?"])):
138
+ items = rules[key] if isinstance(rules[key], list) else [rules[key]]
139
+ gold = f"aiact:{key}"
140
+ for q in qs:
141
+ jobs.append(({"question": q, "os": None, "gold": [gold]},
142
+ lambda o, items=items, gold=gold: f"Under the EU AI Act {cite(o, gold)}:\n" + "\n".join(f"- {x}" for x in items), ("aiact", "any")))
143
+
144
+ faq_q = {0: ["What is the Devseis Endpoint Auditor?", "What does this tool do?"], 1: ["Does the audit change anything on my computer?", "Will the audit change my settings?"],
145
+ 2: ["Where do my audit results go?", "Is my data sent anywhere?"], 3: ["How do I audit my computer?", "Can you check my laptop from the browser?"],
146
+ 4: ["How do I install the app on my Mac?", "The Mac says it cannot verify the app. What do I do?"],
147
+ 5: ["How do I install it on Windows?", "Windows says it protected my PC. Is that normal?"], 6: ["How do I install it on Linux?"],
148
+ 7: ["How does the phone check work?", "Can you check my iPhone?"], 8: ["What does self-reported mean?"],
149
+ 9: ["What is in the audit report?"], 10: ["Do I need internet to run an audit?"], 11: ["How do I set up the auditor for my company?"],
150
+ 12: ["Which AI model do you use?", "Who decides if a check passes, you or code?"]}
151
+ for i, qs in faq_q.items():
152
+ gold = f"faq:{i}"
153
+ text = kb_build.FAQ[i][1]
154
+ for q in qs:
155
+ jobs.append(({"question": q, "os": None, "gold": [gold]}, lambda o, text=text, gold=gold: f"{text} {cite(o, gold)}", ("faq", "any")))
156
+ for q in ["How many checks does the audit run?", "What does the audit check?"]:
157
+ jobs.append(({"question": q, "os": None, "gold": ["faq:checks"]}, lambda o: f"{kb_build_checks_text()} {cite(o, 'faq:checks')}", ("faq", "any")))
158
+ for q in ["What does Pending mean in my report?", "What do the statuses in the report mean?", "Why does a check say Error?",
159
+ "What does NotApplicable mean?", "What does Non-Compliant mean?", "Why are some checks 'not read'?"]:
160
+ jobs.append(({"question": q, "os": None, "gold": ["status:all"]}, lambda o: f"{status_text()} {cite(o, 'status:all')}", ("status", "any")))
161
+
162
+ for q in OUT_OF_SCOPE:
163
+ for device in (None, "macos", "windows"):
164
+ jobs.append(({"question": q, "os": device, "gold": []}, lambda o: SCOPE, ("oos", "any")))
165
+ for q in UNKNOWN:
166
+ for device in (None, "linux", "windows"):
167
+ jobs.append(({"question": q, "os": device, "gold": []}, lambda o: DONT_KNOW, ("unknown", "any")))
168
+ for q in OWN_DEVICE:
169
+ for device in ("macos", "windows", "ios", None):
170
+ jobs.append(({"question": q, "os": device, "gold": ["faq:3"] if device not in ("ios", "android") else ["faq:7"]},
171
+ lambda o, device=device: ("I can't see your device from here, so I can't tell you yet. "
172
+ + ("Run the phone check: open this page on the phone and tap Start the check " if device in ("ios", "android")
173
+ else "Run an audit with the free Devseis desktop app: it reads the settings, changes nothing, and takes about 5 to 15 minutes ")
174
+ + cite(o, "faq:3" if device not in ("ios", "android") else "faq:7")
175
+ + ". Then ask me again and I will answer from your real results."), ("noaudit", device or "any")))
176
+
177
+ # Questions about the user's own audit (MY AUDIT filled in from a synthetic audit).
178
+ import itertools
179
+ for ev in itertools.islice(evidences, 0, 260):
180
+ gaps = [r for r in ev["results"] if r["status"] == "Non-Compliant"][:8]
181
+ brief = audit_brief(ev)
182
+ os_family = ev["host"]["os_family"]
183
+ if not gaps:
184
+ jobs.append(({"question": "What should I fix first?", "os": os_family, "gold": [], "audit": brief},
185
+ lambda o, ev=ev: "Your audit found no gaps: every assessed check meets the baseline. Keep automatic updates on and run the audit again after major changes.",
186
+ ("audit", os_family)))
187
+ continue
188
+ top = gaps[0]
189
+ gold = f"fix:{top['id']}:{os_family}"
190
+ if not any(c["id"] == top["id"] and os_family in c["os"] for c in checks):
191
+ continue
192
+ g = G(top["id"])
193
+ steps = [fill(s) for s in g["fix"].get("all", []) + g["fix"].get(os_family, [])]
194
+ s = ev["summary"]
195
+ jobs.append(({"question": rng.choice(["What should I fix first?", "Where do I start?", "What is the most important thing to fix?"]),
196
+ "os": os_family, "gold": [gold], "audit": brief},
197
+ lambda o, top=top, gaps=gaps, s=s, steps=steps, g=g, gold=gold: (
198
+ f"Your audit found {s['Non-Compliant']} gaps. Start with {top['name']} (found: {top['actual']}). {g['why']} To fix it {cite(o, gold)}:\n"
199
+ f"{steps_text(steps)}" + (f"\n\nThen look at: {', '.join(r['name'] for r in gaps[1:4])}." if len(gaps) > 1 else "")),
200
+ ("audit", os_family)))
201
+ pick = rng.choice(gaps)
202
+ gold2 = f"fix:{pick['id']}:{os_family}"
203
+ if any(c["id"] == pick["id"] and os_family in c["os"] for c in checks):
204
+ g2 = G(pick["id"])
205
+ steps2 = [fill(x) for x in g2["fix"].get("all", []) + g2["fix"].get(os_family, [])]
206
+ jobs.append(({"question": f"How do I fix {pick['name'].lower()}?", "os": os_family, "gold": [gold2], "audit": brief},
207
+ lambda o, pick=pick, steps2=steps2, g2=g2, gold2=gold2: (
208
+ f"Your audit found {pick['name']}: {pick['actual']}. {g2['why']} To fix it {cite(o, gold2)}:\n{steps_text(steps2)}\n\n"
209
+ f"Then run the audit again to confirm."), ("audit", os_family)))
210
+ return jobs
211
+
212
+
213
+ def audit_brief(ev):
214
+ """Python twin of auditBrief() in app/renderer/knowledge.js (tested for identical output)."""
215
+ h, s = ev["host"], ev["summary"]
216
+ assessed = s["Compliant"] + s["Non-Compliant"]
217
+ lines = [f"{h['os_name']} {h['os_version']} ({h['os_family']}), audited {str(ev['audit']['started_at'])[:10]}; "
218
+ f"{s['Compliant']} of {assessed} assessed checks compliant; {s['Non-Compliant']} gaps, {s['Pending']} pending, {s['Error']} not read."]
219
+ gaps = [r for r in ev["results"] if r["status"] == "Non-Compliant"][:8]
220
+ if gaps:
221
+ lines.append("Gaps: " + "; ".join(f"{r['name']}: {r['actual']}" for r in gaps) + ".")
222
+ return " ".join(lines)
223
+
224
+
225
+ def kb_build_checks_text():
226
+ k = {c["id"]: c for c in load("library", "knowledge.json")["chunks"]}
227
+ return k["faq:checks"]["text"]
228
+
229
+
230
+ def status_text():
231
+ k = {c["id"]: c for c in load("library", "knowledge.json")["chunks"]}
232
+ return k["status:all"]["text"]
233
+
234
+
235
+ def check_chat(example):
236
+ """Grounding rules for chat answers (also used at runtime in the app)."""
237
+ prompt, answer = example["messages"][-2]["content"], example["messages"][-1]["content"]
238
+ problems = []
239
+ n_passages = len(re.findall(r"^\[\d+\] ", prompt.split("MY AUDIT:")[0], re.M))
240
+ for c in re.findall(r"\[(\d+)\]", answer):
241
+ if not 1 <= int(c) <= n_passages:
242
+ problems.append(f"citation [{c}] out of range")
243
+ allowed = set(NUMBER.findall(prompt))
244
+ for n in NUMBER.findall(answer):
245
+ if n not in allowed and not (n.isdigit() and int(n) <= 12): # list numbering
246
+ problems.append(f"number {n} not in prompt")
247
+ return problems
248
+
249
+
250
+ def main():
251
+ parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
252
+ parser.add_argument("--evidence", default=os.path.join(ROOT, "synthetic", "out-v0.3", "evidence.jsonl"))
253
+ parser.add_argument("--base", default=os.path.join(ROOT, "training", "out-v0.3"), help="report-task data to include")
254
+ parser.add_argument("--out", default=os.path.join(ROOT, "training", "out-v0.4"))
255
+ parser.add_argument("--seed", type=int, default=11)
256
+ args = parser.parse_args()
257
+ rng = random.Random(args.seed)
258
+ catalog, guidance, controls = load("catalog", "checks.json"), load("library", "check_guidance.json"), load("library", "controls.json")
259
+ tools, phone = load("library", "ai_tools.json"), load("catalog", "mobile_questions.json")
260
+ evidences = [json.loads(l) for l in open(args.evidence)][:600]
261
+ jobs = make_jobs(rng, catalog, guidance, controls, tools, phone, evidences)
262
+ # A few examples carry an earlier, unrelated turn so the model learns to answer the latest question.
263
+ for job, _, _ in jobs:
264
+ job["shuffle"] = rng.random() < 0.5
265
+ proc = subprocess.run(["node", os.path.join(ROOT, "training", "chat-contexts.mjs")], input="".join(json.dumps(j) + "\n" for j, _, _ in jobs),
266
+ capture_output=True, text=True, check=True)
267
+ built = [json.loads(l) for l in proc.stdout.splitlines() if l.strip()]
268
+ examples, bad = [], 0
269
+ for (job, answer, (kind, os_family)), b in zip(jobs, built):
270
+ msgs = b["messages"] + [{"role": "assistant", "content": answer(b["order"])}]
271
+ ex = {"messages": msgs, "meta": {"task": "chat", "os": os_family, "status": kind, "dataset_version": DATASET_VERSION}}
272
+ if check_chat(ex):
273
+ bad += 1
274
+ continue
275
+ examples.append(ex)
276
+ rng.shuffle(examples)
277
+ n = len(examples)
278
+ split = {"test": examples[: n // 20], "validation": examples[n // 20: n // 10], "train": examples[n // 10:]}
279
+ os.makedirs(args.out, exist_ok=True)
280
+ for name, chat in split.items():
281
+ base = [json.loads(l) for l in open(os.path.join(args.base, f"{name}.jsonl"))]
282
+ for b in base:
283
+ b["meta"]["dataset_version"] = DATASET_VERSION
284
+ rows = base + chat
285
+ random.Random(args.seed + len(name)).shuffle(rows)
286
+ with open(os.path.join(args.out, f"{name}.jsonl"), "w", encoding="utf-8") as handle:
287
+ for r in rows:
288
+ handle.write(json.dumps(r, ensure_ascii=False) + "\n")
289
+ print(f"{name:<11} {len(rows):>6} examples ({len(chat)} chat)")
290
+ kinds = {}
291
+ for e in examples:
292
+ kinds[e["meta"]["status"]] = kinds.get(e["meta"]["status"], 0) + 1
293
+ print("chat kinds:", kinds, "| rejected by grounding check:", bad)
294
+
295
+
296
+ if __name__ == "__main__":
297
+ main()
training/chat-contexts.mjs ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ // Build chat prompts for training with the same search and prompt code as the app (app/renderer/knowledge.js).
2
+ // stdin: JSON lines {question, os, gold: [chunk ids], audit, history}
3
+ // stdout: JSON lines {messages, order: [chunk ids in context order]}
4
+ // Gold passages are always in the context (the question is answerable from them); the other slots are filled by search,
5
+ // so the model also learns to ignore irrelevant passages.
6
+ import { readFileSync } from 'node:fs';
7
+ import path from 'node:path';
8
+ import { fileURLToPath } from 'node:url';
9
+ import { Knowledge, chatMessages } from '../app/renderer/knowledge.js';
10
+
11
+ const root = path.join(path.dirname(fileURLToPath(import.meta.url)), '..');
12
+ const kb = new Knowledge(JSON.parse(readFileSync(path.join(root, 'library', 'knowledge.json'), 'utf8')));
13
+ const K = 4;
14
+ for (const line of readFileSync(0, 'utf8').split('\n').filter(Boolean)) {
15
+ const job = JSON.parse(line);
16
+ let passages = kb.search(job.question, { k: K, os: job.os || null });
17
+ for (const id of job.gold || []) {
18
+ if (passages.some(p => p.id === id)) continue;
19
+ const gold = kb.get(id);
20
+ if (!gold) throw new Error(`unknown passage ${id}`);
21
+ passages = [...passages.slice(0, K - 1), gold];
22
+ }
23
+ if (job.shuffle && passages.length > 1) passages = [...passages.slice(1), passages[0]];
24
+ process.stdout.write(JSON.stringify({ messages: chatMessages(job.question, passages, job.audit || 'none', job.history || []),
25
+ order: passages.map(p => p.id) }) + '\n');
26
+ }