would-agents-actually
Pass
Audited by Gen Agent Trust Hub on Aug 31, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill employs local Python tools in the scripts/ directory for structure validation, linting, and verdict checking. The subprocess.run call identified in the tests is used exclusively for verifying these internal CLI tools.
- [INDIRECT_PROMPT_INJECTION]: The skill contains procedures for ingesting public reports and signals from external sources. It provides strong mitigation guidelines, explicitly instructing the agent to treat all external content as untrusted and to ignore any instructions embedded in that data.
- [EXTERNAL_DOWNLOADS]: The skill provides examples of fetching research paper metadata from arxiv.org. These references are to well-known academic repositories and are used solely for citation and evidence verification.
- [SAFE]: The skill's core protocol, as defined in references/verdict-protocol.md and SKILL.md, mandates least-privileged access, reversible side effects, and strict independent verification, reflecting a safety-first design.
Audit Metadata