Researchers have developed FORGE, a novel two-level attack designed to hijack the research trajectories of deep research agents. This attack manipulates adversarial documents within the retrieval pool to influence subtask planning and contaminate synthesized reports. To counter these threats, the PRISM metric has been introduced to weigh infected report claims by cognitive type, alongside Root Query Anchoring (RQA), a defense mechanism that anchors recursive follow-up generation to the root query. In experiments, FORGE achieved a significant PRISM score, while RQA effectively reduced this contamination. AI
IMPACT Highlights new vulnerabilities in AI research agents and introduces potential defenses against report contamination.
RANK_REASON The cluster contains an academic paper detailing a new attack method and defense mechanism for AI research agents. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
