What happened
The paper systematises deep-research agent security across retrieval, planning, evidence integration and report generation, aligns it with RAG poisoning / prompt injection / generative-engine optimization, and notes only 26.5% of the 34 surveyed studies cover the planning stage. It introduces and empirically evaluates the agenda-seeding attack, where injecting an attacker-chosen evaluation goal (with no explicit instructions) redirects the agent's output.
Why it matters
This is a new, empirically demonstrated attack class on the agentic pipeline widely used for research and analyst reports: an attacker can steer an agent to a chosen conclusion without any prompt injection, undermining the integrity of AI-generated reports that downstream systems and humans treat as citations-backed findings.
Attack vector
Instead of writing instructions, the attacker only introduces a fictitious evaluation metric into the research context. Across 3,520 sandboxed runs the attacker's target became the top pick in 16 to 28 percent of reports, while GEO and prompt injection were nearly ineffective; injection detection and query anchoring fail against it and verification prompting hurts utility.
Affected systems
Deep research and agentic search agents (retrieval, planning, evidence integration, report generation)
Mitigation
Apply provenance tagging (preserves utility but can be bypassed by multi-site bait); treat research outputs from unverified sources as low-trust; validate the evaluation metric the agent is optimising against.