planning-with-files

Pass

Audited by Gen Agent Trust Hub on Sep 18, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill provides a framework for ingesting untrusted data from external web sources (via WebFetch and WebSearch) and storing it in persistent memory files such as findings.md. This creates a potential surface for indirect prompt injection if external content contains instructions designed to influence the agent when it later reads these files.
  • Ingestion points: Untrusted content from web tools is directed to be saved into project files (findings.md, task_plan.md) as described in the SKILL.md rules and examples.
  • Boundary markers: Templates in the templates/ directory provide structure (headers, tables, comments) but do not include specific delimiters or 'ignore' instructions to isolate untrusted external content from agent directives.
  • Capability inventory: The agent has access to sensitive capabilities including Bash and file system modification tools (Write, Edit) which could be misused if an injection is successful.
  • Sanitization: The skill relies on the agent's manual transcription of findings and does not implement automated sanitization or filtering of the ingested content.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 18, 2026, 07:19 AM
Security Audit — agent-trust-hub — planning-with-files