planning-with-files

Pass

Audited by Gen Agent Trust Hub on Mar 10, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes lifecycle hooks to run shell commands and internal scripts. It executes cat task_plan.md to provide context during tool use and calls a vendor-provided script ${CLAUDE_PLUGIN_ROOT}/scripts/check-complete.sh when the session terminates.\n- [PROMPT_INJECTION]: The skill's architecture creates an indirect prompt injection surface by integrating external data into the agent's workflow.\n
  • Ingestion points: The agent ingests data from untrusted sources via WebFetch and WebSearch, as well as from persistent local files like task_plan.md and findings.md (SKILL.md).\n
  • Boundary markers: There are no explicit delimiters or warnings instructing the agent to ignore instructions embedded in the planning files or web content.\n
  • Capability inventory: The skill grants access to the Bash and Write tools, which could be leveraged if the agent follows malicious instructions found in processed data.\n
  • Sanitization: The skill does not implement any validation or sanitization for data retrieved through its tools or read from the filesystem.\n- [EXTERNAL_DOWNLOADS]: The skill allows the use of WebFetch and WebSearch tools, enabling the agent to download and process content from external web domains.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 10, 2026, 05:57 PM