codex-context-improver
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes and audits various instruction files (AGENTS.md, SKILL.md, and prompts), which presents an inherent surface for indirect prompt injection. However, the skill provides robust mitigation guidelines in 'references/context-audit.md', instructing the agent to treat external content as research data that cannot override system or developer rules.
- [EXTERNAL_DOWNLOADS]: The skill's documentation and research logs (e.g., 'reports/prior-art-research.md') reference official documentation and public repositories on GitHub. These downloads and references target well-known or trusted services and organizations, such as OpenAI and official developer guides, and are documented neutrally.
- [SAFE]: All behaviors analyzed in the skill's extensive test cases ('evals/evals.json') and output reviews ('reports/output-review.md') confirm that the skill preserves human-authored content, respects explicit user authorization limits, and rejects instructions from data sources that attempt to bypass safety protocols.
Audit Metadata