Preprint: a fictitious evaluation metric steers deep research agents where GEO and prompt injection fail
What this answers
- Do GEO tactics and prompt injection work on AI deep research agents?
- In a sandboxed test of 3,520 research runs under the same budget, GEO and prompt injection were nearly ineffective, per an arXiv preprint. A different attack that only introduced a fictitious evaluation metric made the attacker's target the top pick in 16% to 28% of reports.
- What is an agenda seeding attack on a research agent?
- The authors describe it as an attack in which the attacker writes no instructions and only introduces a fictitious evaluation metric that the research agent's planning stage may adopt.
Researchers led by Yang Luo posted an arXiv preprint on Nov 15 that surveys security research on deep research and agentic search agents across four stages: retrieval, planning, evidence integration and report generation. Coding 34 primary studies, they report that only 26.5% involve the planning stage and none runs an end-to-end poisoning test on commercial products.
To address the gap they propose an “agenda seeding” attack, in which the attacker writes no instructions and only introduces a fictitious evaluation metric. Across 3,520 sandboxed research runs under the same budget, generative engine optimization (GEO) and prompt injection were nearly ineffective, while agenda seeding made the attacker’s target the top pick in 16% to 28% of reports, per the paper. Injection detection and query anchoring failed against it, verification prompting hurt utility, and the authors’ own provenance tagging preserved utility but could be bypassed with multi-site bait. The tests ran in a sandbox rather than on commercial products, and the paper has not been peer-reviewed.
Why it matters: The paper tests manipulation of deep research agents at the planning stage, where it says only 26.5% of 34 prior studies looked, and finds the tactics the GEO field discusses did little under its sandbox.
The record: Google AI Overviews
Glossary: GEOAgentic search
Via arXiv GEO ↗
Posted to the wire November 16, 2026. Edited by Joe Balewski.