self-iteration
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze data from potentially untrusted sources such as command outputs, CI logs, and review findings to generate proposals for changing its own instructions. \n
- Ingestion points: Command outputs, CI logs, review findings, and existing configuration files mentioned in SKILL.md. \n
- Boundary markers: The skill mandates the use of a structured JSON schema (self-iteration-proposal/v1) and a Markdown file located in a specific directory (artifacts/reflections/) for all output. \n
- Capability inventory: The skill utilizes file read/write capabilities (within the artifacts/ path) and expects access to standard utilities like wc for content measurement. \n
- Sanitization: The primary defense is a mandatory human-in-the-loop review process; the agent is explicitly forbidden from applying the changes it proposes. \n- [COMMAND_EXECUTION]: The skill requires the agent to run the wc -c command to verify that proposed changes remain within a budget. This is a standard utility used for legitimate operational constraints.
Audit Metadata