harness-setup

Warn

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill includes 'Codex compatibility notes' that attempt to redefine its own authorization and execution environment. It explicitly instructs the agent that skill activation 'authorizes use of the required spawn_agent subagent(s)', which is a directive to bypass or override platform-level permission controls for subagents. It also tells the agent to execute skill content directly with 'Codex tools' without using standard tool-calling mechanisms, which can be interpreted as a platform behavior override.
  • [INDIRECT_PROMPT_INJECTION]: The skill's core workflow involves reading external project data (plans, architecture reports, tech-stack comparisons) and using this information to generate or modify high-privilege instruction files like CLAUDE.md and AGENTS.md. This creates a vulnerability where malicious content in those external reports could influence the agent's future behavior or instructions.
  • Ingestion points: Reads plan.md, architecture --mode=design output, and tech-stack reports during 'Phase A
  • Stack Detection'.
  • Boundary markers: No explicit markers or 'ignore' instructions are used when reading these external files; the skill relies on a user confirmation step (ask user question tool) which may not detect subtle injections.
  • Capability inventory: The skill has capabilities for file reading/writing, subagent creation via spawn_agent, and the invocation of multiple other complex skills.
  • Sanitization: There is no evidence of sanitization or validation performed on the ingested report data before it is interpolated into output instruction files.
  • [COMMAND_EXECUTION]: The instructions include a requirement to execute a specific script (inject_engineering_foundation_gate.py) at the end of a synchronization process. While this appears to be a local tool, it represents a command execution capability triggered by the skill's logic.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 4, 2026, 10:12 AM