codex-image-gen
Pass
Audited by Gen Agent Trust Hub on Aug 17, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill accesses the sensitive local authentication file
~/.codex/auth.jsonto retrieve and refresh OAuth tokens. These tokens are transmitted exclusively to official OpenAI domains (auth.openai.comandchatgpt.com) to facilitate image generation as requested by the user. The script implements security best practices by applying restrictive file permissions (chmod 600) when updating the credential file.\n- [COMMAND_EXECUTION]: The test suite (tests/test_codex_image.py) usessubprocess.runto execute the main script. This is a benign diagnostic measure used during automated testing to verify CLI help outputs and argument handling.\n- [PROMPT_INJECTION]: The skill ingests untrusted data from user prompts and local files, creating a potential surface for indirect prompt injection.\n - Ingestion points: Inputs are received through the
--prompt,--prompt-file, and--imageCLI arguments inscripts/codex_image.py.\n - Boundary markers: None. User-supplied text is interpolated directly into the JSON request body sent to the backend API.\n
- Capability inventory: The script possesses capabilities for local file system read/write access and network communication with OpenAI APIs.\n
- Sanitization: No local sanitization or filtering is performed on user inputs; the skill relies on the safety guardrails of the remote image generation service.
Audit Metadata