codex-delegation

Pass

Audited by Gen Agent Trust Hub on Sep 29, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The CODEX-INVOCATION.md file provides extensive Bash templates for the agent to execute, including tool invocation (codex), directory management (mktemp), and source control monitoring (git status, git diff).
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to delegate task implementation to a sub-agent (codex), which creates an attack surface where malicious input could influence downstream actions. Ingestion points: User specifications and plans are ingested via stdin to the codex exec command. Boundary markers: The instructions recommend setting explicit stop points and provide code to disable unnecessary integrations. Capability inventory: The sub-agent has workspace-write capabilities and access to configured MCP integrations. Sanitization: The workflow relies on manual review of generated diffs by the human caller.
  • [DATA_EXFILTRATION]: The skill exposes the username 'lecoqjacob' through the hardcoded absolute path /Users/lecoqjacob/.local/bin/codex in the invocation instructions, revealing environment-specific metadata.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 29, 2026, 07:48 PM