baoyu-article-illustrator
Warn
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The system prompt template in
prompts/system.mdcontains explicit instructions intended to bypass AI safety guardrails. It directs the downstream agent to "create stylistically similar alternatives" when encountering sensitive or copyrighted figures and explicitly commands "DO NOT refuse to generate," which is a known pattern for attempting to override safety and refusal policies. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted article content and user-defined metadata which are subsequently interpolated into prompts for an external image generation tool.
- Ingestion points: The skill ingests article files (
path/to/article.md) and direct user content as described inSKILL.mdunder the "Analyze Content" section. - Boundary markers: There are no boundary markers or specialized delimiters used when interpolating the article's core arguments or content into the image generation prompts to prevent the downstream model from following instructions embedded within the article text.
- Capability inventory: The skill has the capability to generate visual descriptions and prompts which are then passed to an image generation tool ("nano banana pro") as specified in
prompts/system.md. - Sanitization: The workflow lacks sanitization or filtering logic to detect or neutralize malicious instructions contained within the articles being illustrated.
Audit Metadata