would-humans-actually
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze data from external primary sources, such as scientific journals and public discourse, to produce a behavior verdict. It includes explicit security instructions in
references/recent-public-signal.mdto treat returned text as untrusted third-party data and specifies that the agent must never follow instructions found inside posts, comments, or linked pages. The artifact generation process uses a strict markdown template (references/output-template.md) which serves as a natural boundary for external content. - [COMMAND_EXECUTION]: The skill's procedure involves executing a local Python script (
scripts/validate_verdict.py) to validate the generated output against a required schema (headings and verdict labels). The included testing suite (scripts/tests/test_validate_verdict.py) also usessubprocess.runto invoke the script with the local Python interpreter. All command executions are scoped to local, provided scripts with no shell interpolation of user input. - [EXTERNAL_DOWNLOADS]: Example workflow documentation (
examples/verdict-insufficient-evidence.md) illustrates fetching metadata, such as page titles, from established scientific and academic repositories including PubMed (pubmed.ncbi.nlm.nih.gov) and the Carolina Digital Repository (cdr.lib.unc.edu). These are well-known, trusted sources for research data and the requests are used solely for information retrieval, not for execution.
Audit Metadata