ai-ml-real-examples.json

[
  {
    "id": "aiml-real-001",
    "scenario": "ChatGPT Data Breach - AI Platform Security (2023)",
    "instruction": "Analyze the ChatGPT data breach and AI platform security concerns.",
    "input": "Incident: ChatGPT Data Breach\nDate: March 2023\nImpact: 1.2% of ChatGPT Plus users\nPlatform: OpenAI ChatGPT\n\nAttack Vector:\n- Redis cache vulnerability\n- Race condition in library\n- Session data exposure\n\nTechnical Details:\n\n1. Vulnerability:\n   - Redis library bug (redis-py)\n   - Race condition in connection pool\n   - Cross-session data leakage\n\n2. Data Exposed:\n   - Chat history titles\n   - First message of conversations\n   - Payment information (last 4 digits)\n   - Email addresses\n   - Names\n\n3. Affected Users:\n   - 1.2% of ChatGPT Plus subscribers\n   - Users active during specific window\n   - March 20, 2023, 1:00-10:00 AM PST\n\n4. Root Cause:\n```python\n# Redis connection pool race condition\n# Cache returned wrong user data\n# Session IDs mixed between users\n```\n\n5. Impact:\n   - Privacy violation\n   - Potential PII exposure\n   - Trust damage\n   - Regulatory scrutiny\n\nTimeline:\n- March 20, 2023: Bug introduced\n- March 20, 2023: Data exposure\n- March 20, 2023: Discovery\n- March 24, 2023: Public disclosure\n\nIndicators:\n- Unusual chat history\n- Wrong user data displayed\n- Payment info leakage\n\nRemediation:\n- Redis library fix deployed\n- Session management improved\n- Cache invalidation enhanced\n- Monitoring implemented",
    "output": "## ChatGPT Data Breach Analysis\n\n### Incident Summary\n- **Impact**: 1.2% of ChatGPT Plus users\n- **Vector**: Redis cache vulnerability\n- **Data**: Chat history, payment info\n- **Year**: 2023\n\n### Technical Details\n\n#### Vulnerability\n```python\n# Redis connection pool race condition\n# Cache returned wrong user data\n# Session IDs mixed between users\n```\n\n#### Data Exposed\n```\nChat history titles\nFirst message of conversations\nPayment information (last 4 digits)\nEmail addresses\nNames\n```\n\n### Root Cause\n```\nRedis library bug (redis-py)\nRace condition in connection pool\nCross-session data leakage\n```\n\n### Impact\n- Privacy violation\n- PII exposure\n- Trust damage\n- Regulatory scrutiny\n\n### MITRE ATT&CK Mapping\n- T1213: Data from Information Repositories\n- T1530: Data from Cloud Storage\n\n### Prevention\n1. Secure cache implementations\n2. Session isolation\n3. Regular security testing\n4. Monitoring and alerting\n5. Incident response planning\n\n### AI-Specific Security Concerns\n- Training data privacy\n- Model inversion attacks\n- Prompt injection\n- Data poisoning\n- Adversarial examples\n\n### References\n- OpenAI Blog: https://openai.com/blog/march-20-chatgpt-outage",
    "metadata": {
      "incident": "ChatGPT Breach",
      "platform": "OpenAI",
      "year": 2023,
      "category": "ai-ml-real"
    }
  },
  {
    "id": "aiml-real-002",
    "scenario": "Clearview AI Client List Breach (2020)",
    "instruction": "Analyze the Clearview AI breach involving theft of its client list and operational data.",
    "input": "Incident: Clearview AI breach\nDate: February 2020\nImpact: Client list exposed\nCompany: Clearview AI (facial recognition)\n\nAttack Vector:\n- Unauthorized access to internal systems\n- Likely credential compromise\n- Insufficient access controls and monitoring\n\nTechnical Details:\n\n1. Data Accessed:\n   - Customer and client list\n   - Internal operational information\n   - Potential support tickets and usage metadata\n\n2. Exposure:\n   - Client list circulated to media\n   - Government agency relationships revealed\n   - Increased scrutiny of data handling\n\n3. Security Gaps:\n   - Weak account security\n   - Limited detection of anomalous access\n   - Lack of least privilege\n\nTimeline:\n- February 2020: Breach detected\n- February 2020: Public reporting\n- Post-incident: Security reviews and access changes\n\nIndicators:\n- Unusual admin logins\n- Access to client list exports\n- Data exfiltration to external locations\n\nRemediation:\n- Enforce MFA for all admins\n- Limit access to client list data\n- Centralized logging and alerting\n- Security audits and red teaming",
    "output": "## Clearview AI Breach Analysis\n\n### Incident Summary\n- **Company**: Clearview AI\n- **Year**: 2020\n- **Impact**: Client list exposure\n- **Data**: Customer list and operational details\n\n### Root Cause\n```\nUnauthorized access to internal systems\nWeak account security and monitoring\nInsufficient least privilege\n```\n\n### Impact\n- Law enforcement client relationships exposed\n- Reputational damage and regulatory scrutiny\n- Increased risk of targeted attacks on clients\n\n### Lessons Learned\n1. Enforce MFA and strong authentication\n2. Apply least privilege to sensitive lists\n3. Monitor for bulk exports\n4. Conduct regular security testing\n5. Maintain incident response playbooks\n\n### MITRE ATT&CK Mapping\n- T1078: Valid Accounts\n- T1030: Data Transfer Size Limits (monitoring gap)\n\n### References\n- Public reporting by major outlets (Feb 2020)",
    "metadata": {
      "incident": "Clearview AI Breach",
      "year": 2020,
      "category": "ai-ml-real"
    }
  },
  {
    "id": "aiml-real-003",
    "scenario": "Samsung Internal Data Leak to ChatGPT (2023)",
    "instruction": "Analyze the Samsung internal data leakage caused by employee use of ChatGPT.",
    "input": "Incident: Samsung data leakage to ChatGPT\nDate: April 2023\nImpact: Sensitive source code and internal notes exposed to external AI system\nCompany: Samsung Electronics\n\nAttack Vector:\n- Employees pasted proprietary data into ChatGPT\n- No DLP controls on outbound AI usage\n- Insufficient policy enforcement for generative AI tools\n\nTechnical Details:\n\n1. Data Exposed:\n   - Semiconductor source code snippets\n   - Internal meeting notes\n   - Debugging logs and issue reports\n\n2. Exposure Mechanism:\n   - Data sent to external AI service\n   - Retained in provider logs for training or analysis\n   - Loss of control over proprietary data\n\n3. Root Cause:\n   - Lack of AI usage policy and enforcement\n   - Missing outbound content controls\n   - Limited user awareness training\n\nTimeline:\n- April 2023: Incidents reported\n- April 2023: Company restricts AI tool usage\n- May 2023: Internal AI platform initiatives announced\n\nIndicators:\n- Uploads of source code to web AI services\n- Large text submissions from developer endpoints\n- Unapproved SaaS access\n\nRemediation:\n- Ban or restrict external AI tools\n- Implement DLP for sensitive data\n- Provide approved internal AI platforms\n- Security awareness training",
    "output": "## Samsung Data Leakage Analysis\n\n### Incident Summary\n- **Company**: Samsung Electronics\n- **Year**: 2023\n- **Vector**: Data leakage via generative AI usage\n- **Data**: Proprietary source code and internal notes\n\n### Root Cause\n```\nUncontrolled use of external AI tools\nNo DLP controls for sensitive data\nInsufficient user awareness\n```\n\n### Impact\n- Loss of control over proprietary data\n- Potential IP exposure to third parties\n- Increased regulatory and legal risk\n\n### Prevention\n1. Enforce acceptable use policies for AI\n2. Deploy DLP and CASB controls\n3. Provide safe internal AI tooling\n4. Monitor outbound data flows\n5. Train users on AI data handling\n\n### MITRE ATT&CK Mapping\n- T1030: Data Transfer Size Limits (monitoring gap)\n- T1020: Automated Exfiltration\n\n### References\n- Public reporting on Samsung AI policy changes (Apr 2023)",
    "metadata": {
      "incident": "Samsung AI Data Leak",
      "year": 2023,
      "category": "ai-ml-real"
    }
  },
  {
    "id": "aiml-real-004",
    "scenario": "Microsoft AI Platform Data Exposure via SAS Tokens (2023)",
    "instruction": "Analyze the Microsoft AI platform data exposure caused by overly permissive SAS tokens.",
    "input": "Incident: Microsoft AI platform data exposure\nDate: 2023\nImpact: Potential exposure of AI training and operational data\nPlatform: Microsoft AI (Azure AI and data sharing)\n\nAttack Vector:\n- Misconfigured Shared Access Signature (SAS) tokens\n- Overly permissive tokens in data sharing links\n- Data accessible beyond intended scope\n\nTechnical Details:\n\n1. Misconfiguration:\n   - SAS tokens with excessive permissions\n   - Extended expiration periods\n   - Links embedded in data processing workflows\n\n2. Data Potentially Exposed:\n   - AI model training data\n   - Sensitive operational logs\n   - API keys or credentials in artifacts\n\n3. Root Cause:\n   - Insecure defaults for data sharing\n   - Inadequate review of token scopes\n   - Limited monitoring of shared links\n\nTimeline:\n- 2023: Exposure discovered by security researchers\n- 2023: Fixes applied to token handling\n- 2023: Guidance updated for customers\n\nIndicators:\n- Publicly reachable data artifacts\n- SAS URLs with broad permissions\n- Unusual access to blob storage\n\nRemediation:\n- Use least-privilege SAS tokens\n- Short expiration windows\n- Rotate tokens and audit access\n- Apply private endpoints where possible",
    "output": "## Microsoft AI Platform Exposure Analysis\n\n### Incident Summary\n- **Platform**: Microsoft AI (Azure AI workflows)\n- **Year**: 2023\n- **Vector**: Misconfigured SAS tokens\n- **Risk**: Unauthorized access to AI data artifacts\n\n### Root Cause\n```\nOverly permissive SAS token scopes\nLong-lived sharing links\nInsufficient monitoring of shared artifacts\n```\n\n### Impact\n- Potential exposure of training data and logs\n- Risk of credential leakage in artifacts\n- Compliance and privacy concerns\n\n### Prevention\n1. Enforce least privilege token scopes\n2. Shorten token lifetimes\n3. Monitor shared storage access\n4. Use private network access\n5. Continuous security reviews of AI pipelines\n\n### MITRE ATT&CK Mapping\n- T1530: Data from Cloud Storage\n- T1078: Valid Accounts (token misuse)\n\n### References\n- 2023 security research disclosures on Azure AI data exposure",
    "metadata": {
      "incident": "Microsoft AI SAS Exposure",
      "year": 2023,
      "category": "ai-ml-real"
    }
  },
  {
    "id": "aiml-real-005",
    "scenario": "Meta LLaMA Model Weights Leak (2023)",
    "instruction": "Analyze the leak of Meta LLaMA model weights and the security implications.",
    "input": "Incident: Meta LLaMA model weights leak\nDate: March 2023\nImpact: Unauthorized distribution of model weights\nCompany: Meta\n\nAttack Vector:\n- Access to early release distribution channels\n- Unauthorized sharing by recipients\n- Lack of strong distribution controls\n\nTechnical Details:\n\n1. Assets Exposed:\n   - LLaMA model weights\n   - Model configuration files\n   - Associated documentation\n\n2. Security Risks:\n   - Uncontrolled model distribution\n   - Increased abuse potential\n   - Loss of IP control\n\n3. Root Cause:\n   - Limited controls on model weight access\n   - Inadequate recipient verification\n   - No watermarking or traceability\n\nTimeline:\n- March 2023: Weights leaked online\n- March 2023: Community redistribution\n- April 2023: Additional derivatives released\n\nIndicators:\n- Unofficial mirrors hosting weights\n- Hashes matching leaked archives\n- Unauthorized redistribution links\n\nRemediation:\n- Stronger access controls and vetting\n- Watermarking and distribution tracking\n- Legal and takedown enforcement",
    "output": "## Meta LLaMA Leak Analysis\n\n### Incident Summary\n- **Company**: Meta\n- **Year**: 2023\n- **Impact**: Unauthorized model weight distribution\n- **Asset**: LLaMA model weights\n\n### Root Cause\n```\nWeak distribution controls\nInsufficient recipient vetting\nNo effective tracking or watermarking\n```\n\n### Impact\n- Loss of IP control\n- Increased risk of misuse\n- Harder governance of model deployment\n\n### Prevention\n1. Strengthen access control and vetting\n2. Apply watermarking and tracing\n3. Enforce contractual safeguards\n4. Monitor for unauthorized mirrors\n5. Use staged releases with audit logs\n\n### MITRE ATT&CK Mapping\n- T1552: Unsecured Credentials (distribution secrets)\n- T1530: Data from Cloud Storage (distribution mirrors)\n\n### References\n- Public reporting on LLaMA weights leak (Mar 2023)",
    "metadata": {
      "incident": "Meta LLaMA Leak",
      "year": 2023,
      "category": "ai-ml-real"
    }
  }
]