self-improving

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill creates a self-modification loop where the agent's future behavior is defined by new instructions generated from external inputs, constituting a surface for indirect prompt injection.\n
  • Ingestion points: The workflow identifies user corrections, verified task failures, and environment discoveries as inputs for creating or refining skills.\n
  • Boundary markers: The agent is instructed to generalize lessons and avoid task-specific details, relying on internal logic rather than strict structural delimiters.\n
  • Capability inventory: The agent uses filesystem write access to the ~/.agents/skills/ directory to persistently store and modify its operational instructions.\n
  • Sanitization: The skill contains explicit prohibitions against including credentials, secrets, or sensitive customer data, and requires validation of generated files.\n- [COMMAND_EXECUTION]: The workflow references the execution of a local utility, skills-ref, to validate the syntax and structure of newly created skill files.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 03:43 AM
Security Audit — agent-trust-hub — self-improving