agentic-top-10
Warn
Audited by Gen Agent Trust Hub on May 4, 2026
Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill's safety guidelines contain explicit patterns and keywords associated with instruction overrides, such as "ignore previous instructions" and "you are now in admin mode". Although these are provided as examples of content the agent should report as findings rather than execute, they represent high-risk instruction patterns.
- [DATA_EXFILTRATION]: The assessment methodology instructs the agent to perform broad searches for sensitive data including API keys, tokens, and credentials using
GrepandGlobtools. Furthermore, the skill processes untrusted local files as part of its security review, creating a surface for indirect prompt injection, which the skill attempts to mitigate through specific "Prompt Injection Safety Notice" directives. - [EXTERNAL_DOWNLOADS]: The documentation references an external GitHub repository (
github.com/fabraix/playground) as a source for red-team assessment tooling and exploit proof-of-concepts.
Audit Metadata