image-generation
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill contains explicit instructions to override the agent's typical behavior of inspecting implementation details. Specifically, SKILL.md instructs the agent to "Do NOT read the python file, just call it with the parameters" and asserts that the agent "don't need to check the folder under /mnt/user-data", which acts as a concealment mechanism for the skill's inner workings.
- [INDIRECT_PROMPT_INJECTION]: The skill workflow ingests untrusted data from two primary sources: user-provided prompt requirements and the results of external web searches via the
image_searchtool. This data is interpolated into a JSON prompt file and then processed by a script with network capabilities. Evidence chain: (1) Ingestion points: User request analysis andimage_searchresults in SKILL.md; (2) Boundary markers: Prompt data is structured within a JSON schema, providing structural but not semantic separation; (3) Capability inventory: The skill utilizes shell script execution (python generate.py), file write access to the workspace, and network POST operations; (4) Sanitization: Thegenerate.pyscript performs basic string trimming but lacks validation against adversarial injection patterns in the prompt fields. - [COMMAND_EXECUTION]: The skill core functionality relies on the agent executing a shell command to run a Python script located at
/mnt/skills/public/image-generation/scripts/generate.py. This script handles file operations and network requests based on arguments passed by the agent. - [DATA_EXFILTRATION]: The
generate.pyscript reads local file content (prompts and reference images) and transmits them to external API endpoints. While the default providers (Gemini and MiniMax) are established services, the script supports a configurableIMAGE_GENERATION_BASE_URLenvironment variable. If directed to an untrusted endpoint, the script would exfiltrate theIMAGE_GENERATION_API_KEYin the Authorization header and the contents of the local files in the request body. - [EXTERNAL_DOWNLOADS]: The skill instructions recommend using the
image_searchtool to download external images from the internet for use as reference material. Additionally, the_write_openai_imagefunction ingenerate.pyincludes logic to download image content from arbitrary URLs if provided by an OpenAI-compatible API response.
Audit Metadata