codex-delegate
Pass
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted content from external repositories that may contain malicious instructions targeting the delegated AI worker.
- Ingestion points: The delegated worker AI is configured to read files from the target repository directory, including specific instruction files such as
AGENTS.md(scripts/codex_job.py). - Boundary markers: The skill prompt uses XML-style tags like
<task>and<delegation_rules>to delineate instructions, which mitigates but does not fully prevent instruction confusion. - Capability inventory: The delegated worker has extensive capabilities including file system modification and shell command execution, especially when configured with the
danger-full-accesssandbox mode. - Sanitization: The skill includes an environment-based Git push block, though it is explicitly documented as a non-robust guardrail.
- [COMMAND_EXECUTION]: The skill uses a Python helper script to execute the
codexCLI tool. - Evidence: The
scripts/codex_job.pyscript utilizessubprocess.Popento executecodexcommands with arguments supplied by the agent, such as directory paths and model configuration. - [DATA_EXFILTRATION]: The skill contains logic to prevent unauthorized data transmission via Git.
- Evidence: The
push_block_envfunction inscripts/codex_job.pyattempts to disable Git pushes by rewriting push URLs in the process environment. The implementation is intended as a safety measure for the worker AI rather than a hardened security boundary.
Audit Metadata