run-codex-adversarial-loop

Pass

Audited by Gen Agent Trust Hub on Jul 27, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes several shell commands for orchestration, including git worktree for isolated environment creation, node for executing the Codex companion script, and ls for locating local plugin assets. In references/codex-and-loop-mechanics.md, the path to codex-companion.mjs is dynamically resolved at runtime from the local plugin cache to interface with the well-known openai-codex service.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and processes external codebases. However, it implements a robust mitigation strategy via its Phase 5 'Verification' stage.
  • Ingestion points: Full committed source tree (SKILL.md, Phase 3).
  • Boundary markers: Includes a shared frame with an 'already-known' ledger and an 'anti-fabrication' rule (references/prompt-templates.md).
  • Capability inventory: Git worktree management, local script execution, and integration branch merging (SKILL.md).
  • Sanitization: Findings are subjected to independent verification by a neutral subagent blind to the original source, followed by orchestrator triage, before any fixes are applied (SKILL.md, Phase 5-6).
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 27, 2026, 12:28 PM
Security Audit — agent-trust-hub — run-codex-adversarial-loop