gemini-image
Pass
Audited by Gen Agent Trust Hub on May 2, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Communicates with
generativelanguage.googleapis.comto send prompts and receive image data. This is a trusted, well-known service for Google AI. - [COMMAND_EXECUTION]: The skill uses shell commands to export user input and process files. There is a potential command injection surface where user-provided prompts are passed into shell variables (
export GEMINI_PROMPT='...'). While single-quoting is suggested, it relies on correct agent behavior to avoid exploitation. - [PROMPT_INJECTION]: The skill processes untrusted user data as image descriptions, creating an indirect prompt injection surface.
- Ingestion points: User-provided image descriptions in
SKILL.md(Step 3). - Boundary markers: Suggested single-quotes in bash exports; no formal delimiters in the prompt content.
- Capability inventory: Network access via
curl, file system read/write, and Python execution. - Sanitization: Relies on quoting instructions for the agent; lacks programmatic sanitization.
Audit Metadata