Researchers propose agenda seeding attack on deep research agents
Researchers systematize the security of deep research agents across retrieval, planning, evidence integration, and report generation, proposing a unified threat model and taxonomy. They introduce the agenda seeding attack, where an attacker introduces a fictitious evaluation metric without writing instructions, successfully manipulating report top picks in up to 28% of sandboxed runs. Additionally, the study evaluates existing defenses and outlines eight new research directions for securing agentic search. Open article →