web-design-guidelines
Fail
Audited by Gen Agent Trust Hub on Aug 26, 2026
Risk Level: HIGHPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [METADATA_POISONING]: The skill's author field is set to 'vercel', while the actual author is 'agisota'. This deceptive labeling impersonates a well-known organization to misrepresent the skill's origin and gain unearned trust.
- [UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill retrieves auditing guidelines from Vercel Labs' GitHub repository. This external source provides the rules applied during the UI review process.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided source files, creating a surface where malicious instructions embedded in those files could influence the agent's behavior.
- Ingestion points: User-provided files or patterns (SKILL.md).
- Boundary markers: Absent; there are no instructions for the agent to use delimiters or ignore instructions within the reviewed content.
- Capability inventory: The agent reads local files and performs network requests to fetch guidelines.
- Sanitization: Absent; the agent is instructed to apply guidelines directly to the content of the read files without prior filtering.
Recommendations
- AI detected serious security threats
Audit Metadata