agent-retro
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze session logs from agent harnesses like Claude Code and Codex. These logs contain full transcripts of previous interactions, providing an attack surface for indirect prompt injection if the processed logs contain adversarial instructions intended to influence the agent during the retrospective process.
- Ingestion points: Session logs from local agent harnesses (SKILL.md, Step 1).
- Boundary markers: No specific delimiters or safety warnings for the log content are defined.
- Capability inventory: The skill proposes fixes including new lint rules, CI jobs, and instruction files based on the log analysis (SKILL.md, Categories section).
- Sanitization: Step 5 mandates redacting secrets from the final report, but there is no specified sanitization or escaping of log content during the analysis phase.
- [DATA_EXPOSURE]: The primary function involves reading local session logs which typically contain source code, environment details, and user-agent conversation history. While the skill instructs the agent to redact secrets in the final report, the underlying process requires the agent to have read access to sensitive local logs.
Audit Metadata