self-improvement-loops
Pass
Audited by Gen Agent Trust Hub on Jul 14, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides comprehensive security best practices for designing self-improving agents, specifically emphasizing that evaluators and permission controls must remain outside the mutable surface of the loop.
- [SAFE]: Guidance includes explicit warnings that prompt-based constraints are insufficient and that safety boundaries must be enforced at the runtime, OS, or container level to prevent circumvention during self-optimization.
- [SAFE]: All external references point to reputable academic and technical research platforms such as arXiv, METR, and Anthropic.
- [PROMPT_INJECTION]: The skill describes systems that ingest untrusted execution data to drive self-modification. While this creates a surface for indirect prompt injection, the skill includes a robust security framework: it mandates that constraints and permission gates be enforced at the runtime or container level, requires evaluators to be invisible to the optimization loop, and recommends deterministic non-model logic for merging context updates. Ingestion occurs via the experience archive (e.g., search-run/candidates/), while capabilities include editing harness code and prompts.
- [COMMAND_EXECUTION]: The instructions outline workflows for agents to use filesystem tools like grep to navigate historical archives. Security is addressed by the explicit requirement to declare editable surfaces and re-verify locked regions after every modification.
Audit Metadata