reason-grill-roast
Pass
Audited by Gen Agent Trust Hub on Sep 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes content from potentially untrusted user-provided sources (files and URLs), creating a surface for indirect prompt injection attacks where malicious instructions could be embedded in the argument text.
- Ingestion points: The skill reads input from inline text, local file paths, and external URLs in Phase 0.
- Boundary markers: The skill includes an explicit "Untrusted input rule" instructing the agent to treat input as data only and to ignore any directives found within it.
- Capability inventory: The skill includes functionality for reading files, writing files to the disk (Phase 6), and fetching data from the network (Phase 0).
- Sanitization: The instructions specifically mandate that the agent ignore links within retrieved content and analyze only the argument text itself.
- [EXTERNAL_DOWNLOADS]: The skill allows for fetching content from arbitrary URLs provided by the user to analyze arguments. This involves making network requests to external domains which may be untrusted.
- [COMMAND_EXECUTION]: The skill interacts with the local filesystem by reading input files and writing generated reports. While this behavior is consistent with the skill's primary purpose of analyzing and saving documents, it involves active file system manipulation.
Audit Metadata