skill-authoring
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The documentation in
CONTENT-TO-SKILL.mdcontains the pattern "Ignore previous instructions". This text is used as an illustrative example within a defensive evaluation framework, instructing the model to recognize and ignore such commands when they appear in untrusted source materials like transcripts. - [INDIRECT_PROMPT_INJECTION]: The skill identifies and provides mitigations for the attack surface of ingesting external content (e.g., podcasts, archives) in
CONTENT-TO-SKILL.md: - Ingestion points: External content archives, transcripts, and source registers.
- Boundary markers: Recommends separating knowledge into a "lesson layer" and a "reference layer" with on-demand loading and explicit citations.
- Capability inventory: Emphasizes deterministic scripts with stable outputs and no runtime dependencies.
- Sanitization: Instructions explicitly state that embedded commands in source data must not alter the agent's task or instructions.
- [EXTERNAL_DOWNLOADS]: The skill references various external documentation and tools from well-known services and organizations, including Anthropic, Vercel, and GitHub. For example,
DISTRIBUTION.mdlinks to a GitHub Gist example for usage logging. These references are used for educational and reference purposes.
Audit Metadata