learning-from-mistakes

Pass

Audited by Gen Agent Trust Hub on Aug 22, 2026

Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses the watch-skill CLI tool (a resource associated with the author 'oxbshw') to manage lessons and display performance statistics. These operations interact with the local directory ~/.watch-skill/ and do not invoke unauthorized network or system commands.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided corrections which are later reused in prompts for similar queries.
  • Ingestion points: The <question>, <the wrong answer>, and <the correction> arguments in the watch-skill lessons add command in SKILL.md.
  • Boundary markers: No specific delimiters or safety warnings are defined for the interpolation of these lessons into future prompts.
  • Capability inventory: The skill is restricted to Bash and Read operations, primarily for interacting with the local watch-skill tool.
  • Sanitization: There is no evidence of automated sanitization or filtering for the user-supplied content before it is stored or reused.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 22, 2026, 12:04 AM
Security Audit — agent-trust-hub — learning-from-mistakes