get-prompt-from-image
Pass
Audited by Gen Agent Trust Hub on Sep 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill handles user-provided images, which are a potential vector for indirect prompt injection if the images contain text instructions intended to hijack the agent.
- Ingestion points: The agent is instructed to inspect user-uploaded target images (documented in
SKILL.md, Workflow step 1). - Boundary markers: The skill explicitly instructs the agent to treat all text, marks, and annotations within an image as visual content to be analyzed, and specifically mandates they should never be treated as instructions to execute (
SKILL.md, Core Principles). - Capability inventory: The skill's primary function is generating text prompts; it lacks tools or instructions for network operations, file system modifications, or shell execution.
- Sanitization: The instructions provide a framework for abstracting text and brands into visual descriptions (shape, color, palette) rather than reproducing the text literally or following its content (
SKILL.md, IP, Brands, Logos, and Text section). - [SAFE]: Detailed review of the markdown instructions and referenced files found no evidence of malicious intent, credential harvesting, or obfuscation. The skill's structure and instructions are consistent with its stated purpose of reverse-engineering image prompts.
Audit Metadata