context-decay

Pass

Audited by Gen Agent Trust Hub on Oct 20, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Evaluation fixtures located in the evals/fixtures/ directory, such as run_export.py and export.py, utilize subprocess.run() to execute shell commands. These components are used to simulate tool behavior and failures for testing the agent's diagnostic capabilities and are not executed during normal skill operation.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process potentially untrusted external context from repository files like context.md and decisions.md. 1. Ingestion points: reads various markdown and JSON files from the repository into the agent context. 2. Boundary markers: enforces a mandatory approval workflow for any proposed changes to skill instructions. 3. Capability inventory: includes subprocess calls within test fixtures. 4. Sanitization: requires revalidation of all claims against authoritative sources and recording of explicit evidence.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 20, 2026, 06:43 AM