skills/promptdriven/pdd/pdd-bug/Gen Agent Trust Hub

pdd-bug

Pass

Audited by Gen Agent Trust Hub on Nov 21, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data, specifically software issue descriptions and user-authored stories. This creates a potential surface for indirect prompt injection where malicious instructions could be embedded in the source data.
  • Ingestion points: The reproduce.md step reads issue descriptions, and the prompt-story.md step incorporates user-provided behavior descriptions.
  • Boundary markers: The skill mandates the use of structured documentation formats and specific markdown headers (e.g., ## Domain decisions, ## Rules to check) which serve as organizational boundaries.
  • Capability inventory: The agent can modify source code, write documentation, and execute a variety of vendor-provided CLI tools (pdd) for testing and simulation.
  • Sanitization: The risk is effectively mitigated by the skill's architecture, which requires human approval at multiple stages (gate: approve, gate: pr) and uses isolated subagents for tasks like review and formal verification to prevent malicious instructions from persisting across the entire workflow.
  • [COMMAND_EXECUTION]: The skill utilizes a specialized command-line toolset (pdd) to automate tasks like linting, testing, and environment simulation. These tools are central to the skill's intended functionality and are used within a controlled workflow.
  • [SAFE_PRACTICES]: The skill instructions enforce several high-standard security practices:
  • Dependency Injection: It explicitly requires injecting ports for external resources (databases, network, files) instead of allowing direct system calls, which limits the agent's reach.
  • Information Disclosure Prevention: The instructions forbid the use of absolute file paths or raw logs in verification reports to prevent leaking environment details.
  • Subagent Isolation: Critical verification and review steps are performed by 'fresh' agents with no access to the previous chat history or builder notes, providing a defense-in-depth against context-based attacks.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 21, 2026, 03:21 AM