canvas-design
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [PROMPT_INJECTION]: The skill includes a 'FINAL STEP' section that uses a simulated conversation history ('The user ALREADY said...') to trick the agent into a refinement loop, overriding its normal assessment of work quality.
- [EXTERNAL_DOWNLOADS]: The instructions direct the agent to 'Download and use whatever fonts are needed,' which could result in the retrieval of untrusted assets from the open web without verification or specific source constraints.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied 'niche references' and conceptual foundations to generate art artifacts, creating a vulnerability surface for embedded instructions.
- Ingestion points: Conceptual DNA and subtle references extracted from user prompts.
- Boundary markers: Absent; instructions encourage deep integration of user-provided concepts.
- Capability inventory: Writing .md, .pdf, and .png files; searching local font directories; fetching external resources.
- Sanitization: Absent; there is no instruction to validate or escape user-supplied conceptual references.
- [METADATA_POISONING]: The 'LICENSE.txt' file contains a copyright notice for 'Anthropic, PBC', which conflicts with the provided author context ('huangzida'), indicating inconsistent or deceptive skill metadata.
Audit Metadata