observe-skill

Pass

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input (task descriptions, success/failure details, and error messages) and writes it to persistent log files in the observations/ directory. These files are explicitly intended for consumption by other skills (such as /amend-skill), creating an indirect injection surface.
  • Ingestion points: Data is collected from user-provided descriptions of tasks and outcomes during the "Collect details" methodology phase in SKILL.md.
  • Boundary markers: The skill uses a structured YAML-like format for the log entries but does not include explicit delimiters or instructions for the agent to treat the interpolated content as untrusted or inert data.
  • Capability inventory: The skill methodology describes performing file system operations, specifically creating directories and writing/appending to files.
  • Sanitization: There is no evidence of sanitization, escaping, or validation of the user-provided strings before they are written to the log files.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 9, 2026, 11:59 AM