concept-to-image
Pass
Audited by Gen Agent Trust Hub on Apr 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill core workflow relies on executing a local Python script
scripts/render_to_image.pyvia shell command to perform image rendering. The script launches a headless Chromium browser using the--no-sandboxflag, which reduces security isolation in the execution environment. - [EXTERNAL_DOWNLOADS]: The skill's documentation and error handling instructions direct the user or agent to download and install external software, specifically the
playwrightpackage and thechromiumbrowser binary from public registries. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It takes untrusted user input (the "concept") and uses it to generate HTML content which is then rendered by a headless browser.
- Ingestion points: User concept prompts processed in
SKILL.md(Step 1). - Boundary markers: None present; the AI is instructed to design HTML directly from user descriptions.
- Capability inventory: Execution of
scripts/render_to_image.py(shell access) and browser rendering via Playwright (network and file system access within the browser context). - Sanitization: No sanitization or validation of the user-provided concept is mentioned before it is used to generate executable HTML/SVG code.
Audit Metadata