claude-video-image
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [COMMAND_EXECUTION]: The skill makes extensive use of the
Bashtool to run local Python scripts (scripts/image_generate.py), manage GPU hardware (nvidia-smi), and process media (ffmpeg). This relies on the security of the local execution environment and the integrity of the referenced scripts. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to take arbitrary user text (prompts) and interpolate them directly into shell command arguments.
- Ingestion points: User-provided strings for "generate image" or "create image" requests are passed to the
--promptargument inSKILL.mdexamples. - Boundary markers: The examples use double quotes for the prompt string, but no explicit instructions are provided to sanitize or escape special shell characters (e.g., backticks, semicolons, or dollar signs).
- Capability inventory: The skill uses
Bash,Read, andWritetools, providing a broad surface for impact if a command injection occurs. - Sanitization: There is no evidence of input validation or sanitization within the provided instructions.
- [METADATA_POISONING]: The skill description contains references to non-existent or futuristic AI models and specific pricing models (e.g., "Gemini 3 Pro Image", "Nano Banana Pro", "OpenAI GPT Image 1 Mini"). While these may be flavor text or placeholders, they represent inaccurate metadata that could mislead users or agents regarding the skill's actual capabilities and technical dependencies.
Audit Metadata