deep-research
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large volumes of external data (academic papers, web pages, and API results). It proactively mitigates the risk of indirect prompt injection by including explicit boundary markers (e.g.,
<!-- canonical:instruction-data-boundary -->) and repeated instructions to the agents to treat external content as untrusted data rather than instructions. - [COMMAND_EXECUTION]: Several agents within the pipeline are instructed to execute local Python scripts (e.g.,
scripts/run_ledger.py,scripts/check_pipeline_integrity.py) and standard tools likepdftotextfor data management and verification tasks. These operations are scoped to the skill's legitimate functions and include internal integrity checks. - [EXTERNAL_DOWNLOADS]: The skill performs network requests to well-known academic registries and services (including Crossref, arXiv, OpenAlex, Semantic Scholar, and PubMed) to verify the existence and accuracy of citations. The use of these services is consistent with the skill's stated purpose of academic research.
- [SAFE]: The skill includes robust internal security logic, such as a 'Standalone-Mode Self-Gate' in the report compiler to prevent the bypass of citation locators and strict write-scope guards to maintain phase boundaries between agents.
Audit Metadata