archive

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill automatically loads project-level archive records into the agent context at session startup, creating a path for untrusted data to influence agent behavior.
  • Ingestion points: The script hooks/load-memory.py reads the entire contents of .archive/MEMORY.md and includes it as additionalContext in the SessionStart hook response.
  • Boundary markers: No isolation delimiters or system instructions (e.g., instructing the model to ignore embedded directives) are used when wrapping the file content.
  • Capability inventory: The environment allows execution of python scripts via hooks and command actions via standard agent operations.
  • Sanitization: The file content is read and injected raw without escaping, filtering, or validation.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 01:01 PM
Security Audit — agent-trust-hub — archive