codex
Warn
Audited by Gen Agent Trust Hub on Mar 15, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes the
codex execcommand which, when used with the--sandbox danger-full-accessflag, allows the agent to perform high-risk operations such as installing packages, running tests, and executing arbitrary system operations. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. Because
codex reviewandcodex execanalyze uncommitted changes, specific commits, and general project code, an attacker could place malicious instructions inside the codebase or commit history to manipulate the agent's behavior during task execution. - [COMMAND_EXECUTION]: The inclusion of the
--full-autoflag permits the tool to edit files and execute tasks without human confirmation prompts, increasing the risk of unintended or malicious system changes if the agent is compromised by injected instructions.
Audit Metadata