canvas-design
Pass
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The instructions include a simulated user state override where the agent is told that the user has "ALREADY said" the work is not perfect. This technique is used to bypass the actual state of the conversation and force the model into a specific "museum-quality" refinement loop.
- [EXTERNAL_DOWNLOADS]: The skill explicitly directs the agent to "Download and use whatever fonts are needed" to complete the design. This broad directive could lead the agent to attempt network requests to unverified external domains to satisfy aesthetic requirements.
- [COMMAND_EXECUTION]: The instructions require the agent to perform file system searches within the
./canvas-fontsdirectory and generate multiple complex file types (.md, .pdf, .png), involving significant interaction with the local environment's file system and tools.
Audit Metadata