sports-extreme-stunts
Warn
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONOBFUSCATIONPROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill requires the agent to ingest external reference data from references/cases.md that contains instructions designed to override the agent's persona. The 'E6' case includes a prompt excerpt wrapped in tags that instruct the agent to 'Act as my AI animation director, scriptwriter, illustrator, and production engineer' and use tools like 'Python, HTML Canvas, SVG, FFmpeg'. The workflow in SKILL.md (Step 3) explicitly directs the agent to 'Inspect its... prompt excerpt', ensuring the agent processes these instructions. There is no sanitization or clear boundary instruction to prevent the agent from adopting the roles defined in the reference data.
- [OBFUSCATION]: Static analysis detected hidden Unicode characters (steganography) in references/cases.md. These characters are used to embed instructions that are invisible to human users but processed by the AI model, a technique frequently used to hide malicious logic or prompt injections.
- [PROMPT_INJECTION]: The reference material includes language characteristic of jailbreaks and unauthorized role-play. The 'E6' case uses phrases such as 'NO F*CKING WAY...' and 'steal the prompt before everyone catches on' to steer the agent towards following its embedded instructions.
- [DYNAMIC_EXECUTION]: The prompt excerpt in Case E6 instructs the agent to use execution environments like Python, HTML Canvas, and FFmpeg. If the agent acts on these instructions, it would be performing dynamic code execution based on untrusted data.
- [EXTERNAL_DOWNLOADS]: The skill instructions and reference data direct the agent to fetch media from external domains. The workflow requires the agent to 'Inspect its finished media' using URLs pointing to media.goodcase.ai and goodcase.ai, which are unknown external sources.
Audit Metadata