canvas-design
Pass
Audited by Gen Agent Trust Hub on Feb 27, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSNO_CODE
Full Analysis
- [PROMPT_INJECTION]: The skill employs strong instructional framing and simulated user feedback to override standard agent behavior.
- Evidence: The 'FINAL STEP' section includes a pre-defined user quote ('It isn't perfect enough...') designed to force the model into a specific 'refinement' state, bypassing the actual real-time user interaction.
- Evidence: Instructions use frequent 'CRITICAL' and 'MUST' directives to enforce a specific 'master-level' persona and craftsmanship style.
- Evidence: The prompt explicitly directs the agent to 'STOP' and override its own instincts to call functions or draw new shapes, targeting the model's tool-use logic.
- [EXTERNAL_DOWNLOADS]: The skill instructs the agent to perform external downloads without providing safe or verified sources.
- Evidence: SKILL.md directs the agent to 'Download and use whatever fonts are needed,' which may lead the agent to interact with untrusted third-party domains to find assets.
- [NO_CODE]: The provided skill package contains only markdown instructions and text-based license files, with no functional code or scripts included.
Audit Metadata