ai-act-incidents

/ai-act-incidents

Ground scanner gaps in documented incident history. Returns real-world and research-demonstrated incidents that map to the requested dimension, article, or threat category, with cross-framework taxonomy mappings (OWASP LLM Top 10 2025, OWASP Agentic AI Top 10, NIST AI RMF, MITRE ATLAS) and published mitigations.

Arguments

• $1 (required) — lookup key. Three accepted forms:
  • Dimension id — e.g. security, risk_mgmt, human_oversight, logging, lethal_trifecta
  • Article id — e.g. art9, art15, art72. Accepts the same normalisation as /ai-act-article: art9, Art. 9, 9, ART9.
  • Threat-category id — e.g. prompt_injection, data_exfiltration, model_inversion, privilege_escalation
• --limit N (optional) — number of incidents to return. Defaults to 5.
• --real-world-only (optional) — restrict to corpus="real-world" incidents (excludes research-demonstrated).

Behaviour

1. Normalise $1 to determine lookup type (dimension / article / threat).

2. Call the appropriate corpus API:

   from scanner import incidents_for_dimension, incidents_for_article
   from scanner.incident_grounding import incidents_for_threat
   
   # dimension
   incidents = incidents_for_dimension("security", limit=5)
   
   # article
   incidents = incidents_for_article("art15", limit=5)
   
   # threat
   incidents = incidents_for_threat("prompt_injection", limit=5)
   

3. Present results in this structure:

   # Incidents for: security (Art. 15 — accuracy, robustness, cybersecurity)
   
   Dataset: GenAI & Agentic AI Security Incidents (emmanuelgjr/genai-incidents, CC-BY-4.0)
   Showing 5 of N matching reviewed incidents.
   
   ## INC-03441 — LLM agent exfiltrates CRM records via prompt injection (2024)
   Corpus: real-world | Severity: critical
   OWASP LLM: LLM01, LLM06 | OWASP ASI: ASI01
   MITRE ATLAS: AML.T0051 (LLM Prompt Injection), AML.T0054 (LLM Jailbreak)
   NIST AI RMF: MEASURE-2.7, MANAGE-4.1
   
   Description: [one-sentence summary from incident.description]
   
   Mitigations:
   - Insert a HITL gate before write actions on sensitive data stores
   - Sanitize and validate all inputs from untrusted channels before passing to the LLM
   - Apply output filtering to detect and block exfiltration patterns
   
   EU AI Act relevance:
   - Art. 15(4): prompt-injection resilience is an explicit cybersecurity obligation
   - Art. 9(2)(b): this incident class constitutes foreseeable misuse for any agent with
     external-input access and data-store read capability