keeping-skills-current

Pass

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: SAFEPROMPT_INJECTIONNO_CODE
Full Analysis
  • [PROMPT_INJECTION]: The skill instructs the agent to proactively modify its own core instruction files (SKILL.md) based on observations made during runtime. This creates an indirect prompt injection surface where untrusted data from a task environment could be permanently baked into the agent's instructions.
  • Ingestion points: The agent is directed to monitor for divergences between its documentation and "reality" (tool outputs, command results, and external techniques found in files).
  • Boundary markers: The skill lacks explicit boundaries or warnings to prevent the agent from adopting malicious instructions encountered during a task into its permanent skill set.
  • Capability inventory: The skill explicitly grants and encourages the agent to use file-writing capabilities to update its configuration at ~/.claude/skills/<name>/SKILL.md.
  • Sanitization: There are no instructions provided to validate, sanitize, or escape content derived from the environment before it is written to the instruction files.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 11, 2026, 09:44 PM
Security Audit — agent-trust-hub — keeping-skills-current