canvas-design
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided aesthetic instructions ("subtle input or instructions") to generate design philosophies and visual outputs. It lacks explicit boundary markers or sanitization for this untrusted data. Given the agent's capability to write files (.pdf, .png, .md) and refine code for image generation, this creates a surface where malicious user input could potentially influence the resulting code or content.
- [EXTERNAL_DOWNLOADS]: The instructions explicitly direct the agent to "Download and use whatever fonts are needed," which involves network operations to fetch external resources from unspecified and potentially untrusted third-party sources.
- [PROMPT_INJECTION]: The skill employs a narrative technique that spoofs user history ("The user ALREADY said 'It isn't perfect enough...' ") to override the agent's state and force a refinement process. This style of instruction spoofing is a pattern used to control and manipulate agent behavior.
- [DYNAMIC_EXECUTION]: The skill instructs the agent to "Go back to the code and refine/polish further," which involves the generation and subsequent modification of code (likely Python) to produce the visual masterpieces described in the instructions.
Audit Metadata