planning-with-files

Warn

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: MEDIUMPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (MEDIUM): The skill creates a feedback loop where untrusted data from external sources is stored in persistent files and then used to influence agent reasoning.
  • Ingestion points: Untrusted content enters via research findings written to notes.md and error logs written to task_plan.md.
  • Boundary markers: Absent. The templates do not use delimiters or instructions to treat the file content as untrusted data.
  • Capability inventory: The agent is explicitly instructed to 'Read task_plan.md' and 'refresh goals' before every major action, meaning content in these files directly influences tool selection and execution.
  • Sanitization: None. Malicious instructions contained in a webpage or a tool error could be persisted into the agent's 'working memory' and executed during the next decision loop.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 16, 2026, 09:17 AM