visualize
Pass
Audited by Gen Agent Trust Hub on Aug 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by generating interactive HTML and JavaScript fragments that can trigger agent actions via the
window.openai.sendFollowUpMessageAPI. - Ingestion points: User-provided data and requests are directly processed to generate HTML content and interactive logic as defined in
SKILL.md. - Boundary markers: The skill instructions lack requirements for the agent to sanitize or delimit user data when interpolating it into JavaScript variables or prompt strings used for agent interaction.
- Capability inventory: The skill facilitates local script execution via
scripts/render.pyand allows the generated frontend to initiate follow-up messages that the agent processes. - Sanitization: No explicit validation, filtering, or escaping of external content is mandated before it is embedded in the visualization fragments.
- [COMMAND_EXECUTION]: The skill uses a local Python utility to wrap, export, and serve the generated visualization fragments.
- Evidence:
SKILL.mdcontains instructions to executepython3 scripts/render.py <absolute-fragment-path> [<destination>.html] [--serve]. - [EXTERNAL_DOWNLOADS]: The skill is configured to fetch assets and dependencies from established and well-known Content Delivery Networks (CDNs).
- Evidence: The whitelisted domains include
cdnjs.cloudflare.com,esm.sh,cdn.jsdelivr.net,unpkg.com,fonts.googleapis.com,fonts.gstatic.com, andfonts.bunny.net. - [PROMPT_INJECTION]: The skill metadata and default prompt include instructions to ensure the agent does not refuse requests and describes the capability as unrestricted.
- Evidence: The
openai.yamlfile defines the display name as "Visualize (Unrestricted)" and states the skill has "no scope/stealth gates" in its description.
Audit Metadata