delegated-execution

Pass

Audited by Gen Agent Trust Hub on Apr 5, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by reading task specifications from external plan files and passing them to subagents. \n
  • Ingestion points: SKILL.md describes reading implementation plans from docs/plans/feature-plan.md and extracting task text. \n
  • Boundary markers: The implementer-prompt.md uses headers like ## Task Specification to delimit content, but does not include explicit instructions to ignore embedded agent commands. \n
  • Capability inventory: The builder subagents are authorized to write code, execute shell commands for testing, and commit changes via git. \n
  • Sanitization: No sanitization or filtering of the task text is performed before it is interpolated into the subagent's prompt.\n- [COMMAND_EXECUTION]: The workflow involves subagents performing actions that require shell access. \n
  • Evidence: The implementer-prompt.md template explicitly directs the subagent to 'Write tests', 'Verify the implementation works', and 'Commit your work', which necessitates the execution of development tools and scripts.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 5, 2026, 07:45 PM