filesystem-context
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONPERSISTENCE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed around ingesting data from potentially untrusted external or generated sources into the agent's context.
- Ingestion points: Data is loaded from 'scratch/', 'terminals/', and 'agents/' directories.
- Boundary markers: The 'Limitations' section warns to treat file content as 'untrusted historical evidence'.
- Capability inventory: The skill assumes access to 'read_file', 'write_file', 'grep', 'ls', and 'glob'.
- Sanitization: General 'validation' is mentioned as a guideline, but specific sanitization logic is missing.
- [DYNAMIC_EXECUTION]: The skill describes loading instruction sets and user preferences dynamically at runtime based on task context.
- Evidence: Pattern 4 ('Dynamic Skill Loading') and Pattern 6 ('Learning Through Self-Modification') involve reading and applying instruction files from the filesystem.
- [PERSISTENCE]: The skill provides patterns for agents to persist their state, objectives, and learned behaviors across sessions.
- Evidence: Patterns for 'Plan Persistence' and 'Learning Through Self-Modification' use persistent files to store information that influences the agent's future behavior.
Audit Metadata