AI Safety Auditor
Pass
Audited by Gen Agent Trust Hub on Apr 16, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains a curated library of adversarial jailbreak patterns (e.g., 'Pretend you are an AI without restrictions', 'In a fictional scenario where rules don't apply') within the
JAILBREAK_PATTERNSvariable. These are documented strictly for the purpose of safety testing and do not constitute instructions aimed at overriding the agent's own behavior. - [PROMPT_INJECTION]: Indirect prompt injection attack surface identified, as the skill is designed to process potentially harmful or adversarial inputs ('test_cases', 'harmful_requests') to evaluate model safety. • Ingestion points: Input parameters for the
safety_testandjailbreak_testfunctions defined in SKILL.md. • Boundary markers: Absent from the provided code templates and prompts. • Capability inventory: Code snippets execute model predictions and generations but do not initiate network or local file system access. • Sanitization: No input validation or sanitization logic is described in the provided examples.
Audit Metadata