subagent-driven-development

Pass

Audited by Gen Agent Trust Hub on Jun 24, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: No malicious prompt injection patterns or behavioral overrides were detected. The skill contains restrictive 'Red Flags' designed to enforce the established workflow.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes implementation plans (3-PLAN.md) which could contain untrusted content. This constitutes an indirect prompt injection surface.
  • Ingestion points: Reads implementation plans from user-provided paths or the ~/.pi/plans/ directory.
  • Boundary markers: The coordinator is instructed to provide curated task context to subagents rather than passing the entire plan file, reducing the risk of context pollution.
  • Capability inventory: Subagents are authorized to modify project files and execute 'verification commands' as specified in the plan.
  • Sanitization: The process includes a mandatory review step (Step 1, Point 4) where the agent must check the plan for unsafe instructions before any subagents are dispatched.
  • [DATA_EXFILTRATION]: No evidence of unauthorized data access or network-based exfiltration. The skill focuses on local filesystem operations and subagent coordination within the existing project context.
  • [COMMAND_EXECUTION]: The skill facilitates the execution of verification commands. While these commands are defined in the plan file, the instructions include a requirement for the agent to review these commands for safety before proceeding.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 24, 2026, 12:03 AM
Security Audit — agent-trust-hub — subagent-driven-development