codex-cto
Fail
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: HIGHCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script
scripts/cto-invoke.shis vulnerable to command injection through the use of theevalcommand on a string constructed with unvalidated input. The$PROMPTand$MODELvariables are interpolated directly into a command string that is then executed. An attacker or a malicious prompt could include shell metacharacters (e.g.,;,&,|,\,$()`) to execute arbitrary system commands. - [DATA_EXFILTRATION]: The command injection vulnerability in
cto-invoke.shcan be leveraged to exfiltrate sensitive data. Since the script is executed in the user's environment, a malicious payload could read environment variables, SSH keys, or cloud credentials and transmit them to an external endpoint. - [INDIRECT_PROMPT_INJECTION]: The skill architecture relies on a plan-execute-review loop where an external AI (Codex) provides instructions that the agent executes using powerful native tools. This creates a significant surface for indirect prompt injection.
- Ingestion points: The agent ingests and parses JSON output generated by the
cto-invoke.shscript (documented inSKILL.md). - Boundary markers: There are no protective delimiters or instructions for the agent to ignore potentially malicious commands embedded in the external AI's response.
- Capability inventory: The agent is explicitly granted access to high-privilege tools such as
Bash,Write, andEditto perform the tasks (documented inSKILL.md). - Sanitization: The instructions lack any sanitization or validation logic for the tasks provided by the external service. The directive "Do not deviate from the plan without reason" strongly encourages the agent to follow instructions that could be malicious.
Recommendations
- AI detected serious security threats
Audit Metadata