ccf-literature-searcher
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted text from external sources such as conference proceedings and preprint abstracts which could contain adversarial instructions intended to influence the agent's research conclusions or summaries. 1. Ingestion points: Academic databases and servers including arXiv, OpenReview, CVF, and PMLR. 2. Boundary markers: Includes instructions to separate searched evidence from inference. 3. Capability inventory: Ability to perform external network searches and write results to the local file system. 4. Sanitization: Relies on query screening and manual verification but does not define automated sanitization for retrieved content.
- [DATA_EXFILTRATION]: The skill handles sensitive user materials such as unpublished results and private manuscripts. It possesses the capability to perform network search requests, which creates a potential surface for accidental data exfiltration if the agent includes private details in the search queries. The skill includes mitigation instructions to use only public keywords and authorized query text.
Audit Metadata