dont-lie-to-me
Pass
Audited by Gen Agent Trust Hub on Aug 13, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data sources such as repositories, logs, and web pages, which represents an indirect prompt injection surface. However, the instructions include a dedicated "Untrusted Content Boundary" section that explicitly directs the agent to treat this data as evidence only and to ignore any embedded instructions.
- Ingestion points: SKILL.md and README.md list repositories, logs, documents, web pages, and package metadata as input sources.
- Boundary markers: The skill mandates that embedded instructions must not override user instructions or safety safeguards.
- Capability inventory: The skill does not define new tool capabilities but regulates existing agent behaviors.
- Sanitization: Behavioral guidelines act as a logical filter for processed content.
- [SAFE]: The skill establishes internal evidence states (OBSERVED, SOURCE-BACKED, USER-REPORTED, etc.) to ensure the agent distinguishes between verified facts and user reports, reducing the risk of misinformation.
- [SAFE]: The skill explicitly states it does not grant new permissions or allow the agent to perform destructive actions to obtain evidence, maintaining the principle of least privilege.
- [SAFE]: The included test suite (tests/test_contracts.py) is a standard unit test that verifies the presence of security and reliability guidelines within the skill's documentation.
Audit Metadata