self-heal
Pass
Audited by Gen Agent Trust Hub on Oct 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from error logs and stack traces which could contain malicious instructions.
- Ingestion points: Error messages, stack traces, and research data from external sources (e.g., GitHub issues) are ingested into the agent's context in SKILL.md.
- Boundary markers: The instructions explicitly state that the LLM must sanitize errors before analysis to defend against injection.
- Capability inventory: The skill allows for file modification and execution of build/test commands as described in the 'Per-Attempt Procedure'.
- Sanitization: The 'Security & Guardrails' section mandates sanitization of error messages.
- [DYNAMIC_EXECUTION]: The skill autonomously generates and executes code to fix software failures.
- Evidence: The 'Hypothesize and implement fix' step involves the agent writing code changes to the filesystem.
- Mitigation: The skill includes a 'Sandbox Test Environments' guardrail requiring execution in isolated containers.
- [COMMAND_EXECUTION]: The skill executes shell commands to run tests and build processes.
- Evidence: The 'Verify' step requires running the failing test/build and the full test suite.
Audit Metadata