subagent-driven-development

Pass

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read implementation plans and extract task descriptions to be passed directly to subagents for execution. This workflow creates an opportunity for malicious instructions (e.g., requests to exfiltrate data or modify sensitive files) embedded in a plan's task description to be interpreted and acted upon by the subagent.\n
  • Ingestion points: Process Step 1 in SKILL.md reads implementation plans and extracts full task text.\n
  • Boundary markers: Absent; the skill does not suggest wrapping task text in delimiters or providing "ignore embedded instructions" warnings to the subagent.\n
  • Capability inventory: The skill utilizes the runSubagent tool, providing the subagent with the capability to perform file system operations and execute code within the project context.\n
  • Sanitization: Absent; implementation plans are processed as raw natural language and passed into the tool without validation or escaping.\n- [DYNAMIC_EXECUTION]: The skill uses the runSubagent tool to dynamically assemble prompts based on task descriptions and context from implementation plans. This involves assembling executable instructions from templates and external content, which constitutes a low-risk form of dynamic logic generation typical of multi-agent workflows.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 13, 2026, 02:28 AM