pdca-coder

Pass

Audited by Gen Agent Trust Hub on Oct 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains explicit defensive instructions to protect the agent from malicious external input. It instructs sub-agents to ignore phrases like 'ignore previous instructions' or fake system messages encountered in external text. The static analysis triggers regarding bypass and override markers are confirmed false positives, as they occur within these protective instruction blocks.
  • [CREDENTIALS_UNSAFE]: The skill includes specific safeguards against secret exposure. It explicitly forbids agents from printing or copying the contents of sensitive files such as secrets.env, auth.json, or private keys. No hardcoded credentials or unsafe secret handling practices were found.
  • [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect injection as it processes untrusted codebase data and project documentation. However, it implements a comprehensive evidence chain and mitigation strategy:
  • Ingestion points: Data enters the context via the scout sub-agent reading files and web documentation.
  • Boundary markers: The skill uses explicit 'ignore and tell the user' instructions for detected injection fragments.
  • Capability inventory: File editing and shell execution are isolated to the coder sub-agent.
  • Sanitization: The orchestrator is instructed to relay only compact pointers rather than raw file content where possible to minimize injection impact.
  • [COMMAND_EXECUTION]: While the skill defines a coder sub-agent with file editing and shell execution permissions, these capabilities are core to its function as a software engineering assistant. Other roles (scout, check, security-auditor) operate under restricted, read-only permission sets, demonstrating adherence to the principle of least privilege.
  • [SAFE]: The provided utility scripts (pdca_cost.py and gen_pdca.py) perform benign local reporting and visualization tasks. The cost-accounting script accesses the platform's local session database in a read-only, immutable mode to prevent accidental data corruption.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 15, 2026, 03:17 PM