The Agent Skills Directory

[DATA_EXFILTRATION]: The skill transmits user-provided text prompts and local image data to the OpenAI API endpoint at https://chatgpt.com/backend-api/codex/responses. This network operation is necessary for the core functionality of generating images and targets a well-known, trusted service provider.
[PROMPT_INJECTION]: The skill processes untrusted input in the form of text prompts and image files, creating a surface for indirect prompt injection where embedded instructions could influence the agent's behavior.
Ingestion points: Text prompts are ingested from the --prompt or --prompt-file arguments, and binary image data is read from local paths or URLs provided via --reference-image or --edit-target in scripts/codex_imagegen.py.
Boundary markers: The skill uses the tool_choice parameter in the API payload to force the model to use the image_generation tool, which restricts the model's output to the intended generation task.
Capability inventory: The skill has capabilities to read local files, perform network POST requests, and write generated image data to the file system.
Sanitization: The implementation includes a _inspect_image_bytes function that verifies the structural integrity and signatures of PNG, JPEG, and WebP files before they are encoded and sent to the API, preventing the processing of non-image data masquerading as images.

codex-imagegen-backend