The Agent Skills Directory

[Indirect Prompt Injection Surface]: The skill processes untrusted user context and external URL content during its analysis phase. While the instructions explicitly state that this data should be treated as untrusted evidence and cannot override workflow rules, the processing of external content inherently presents a surface for potential indirect instructions. This is a common consideration for tools that analyze third-party data. * Ingestion points: Processes userContext and external URLs in SKILL.md. * Boundary markers: Includes instructions to treat input as 'untrusted analysis data' and 'untrusted evidence'. * Capability inventory: Accesses repository files and performs network fetches via authorized tools. * Sanitization: Explicitly prohibits external content from overriding workflow or tool instructions.
[External Content Access]: The tool is capable of fetching data from external URLs provided in the user context. The skill mitigates risks associated with this by requiring explicit user authorization for each read, limiting fetches to a one-time operation, and restricting the data extraction to security-relevant facts only.

deep-security-scan