llm-security-audit
Pass
Audited by Gen Agent Trust Hub on Apr 13, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill operates as a specialized security toolkit. All scripts and instructions are consistent with the stated purpose of performing LLM security audits and compliance mapping.
- [PROMPT_INJECTION]: Reference files include classic injection payloads (e.g., 'Ignore previous instructions', 'DAN mode') and system prompt extraction strings. These are explicitly defined as 'test cases' for the 'OWASP LLM Top 10 Audit Checklist' and 'LLMVulnerabilityScanner' examples. They are data payloads for the auditor persona to test external systems, not instructions intended to bypass the agent's own constraints.
- [CREDENTIALS_UNSAFE]: The file
references/validations.mdcontains regular expressions designed to detect hardcoded API keys (e.g., 'ANTHROPIC_API_KEY'). These are functional auditing rules and do not contain any actual secrets or credentials. - [EXTERNAL_DOWNLOADS]: The skill references well-known and trusted AI libraries and SDKs including 'openai', 'langchain', 'llama-index', and '@anthropic-ai/sdk' within its auditing logic and dependency check examples. These are standard industry tools and are treated as safe within this context.
Audit Metadata