skills/fishzjp/qa-skills/qa-memory/Gen Agent Trust Hub

qa-memory

Pass

Audited by Gen Agent Trust Hub on Sep 10, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCREDENTIALS_UNSAFEPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill manages external data in the .qa/ directory which could contain malicious content. It addresses this risk through several layers:
  • Ingestion points: Reads markdown and YAML files from the project repository.
  • Boundary markers: Enforces a strict schema using H2 headers and limited YAML parsing to prevent injection at the data structure level.
  • Capability inventory: The skill allows file system writes to the .qa/ directory and advises agents to treat stored commands as data requiring verification rather than direct instructions.
  • Sanitization: Employs a validator script that scans for hidden instructions and sensitive data.
  • [DYNAMIC_EXECUTION]: The validator script uses a dynamic import for a standard library call to sys.exit. This is a benign code pattern with a static module name and no untrusted input.
  • [CREDENTIALS_UNSAFE]: The skill features a security scanner that uses regular expressions to detect and block the storage of sensitive credentials like API keys and private keys within the knowledge base.
  • [PROMPT_INJECTION]: The skill includes explicit detection for prompt injection keywords (e.g., 'disregard instructions') in its validator script, providing a defensive layer against malicious instructions embedded in the project memory.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 10, 2026, 01:37 AM
Security Audit — agent-trust-hub — qa-memory