get-research-paper
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources (paper titles and abstracts from APIs such as arXiv and PubMed) to generate curated reading lists and синтез briefings. This creates a surface for indirect prompt injection if an attacker-controlled abstract contains instructions intended to influence the agent's behavior.
- Ingestion points: Untrusted external content is ingested during Phase 2 (Source dispatch) as described in
workflows/search.md. - Boundary markers: The skill does not currently implement structural boundary markers or specific delimiters to isolate the external content within the prompts.
- Capability inventory: The skill utilizes
filesystem-writefor generating local artifacts andshell-exec/python3to run its internaltoolchains/arxiv_search.pyscript. - Sanitization: Sanitization is handled via prompt instructions in
prompts/summarization.mdthat direct the model to stay strictly within the source abstract and avoid meta-commentary, which serves as a primary defense.
Audit Metadata