Study: deep research agents vulnerable to citation-free manipulation
A planning-stage exploit called agenda seeding shows AI research agents can be redirected without a single injected instruction, raising the stakes for institutions whose authority depends on correct citation.
Key takeaways
- Zero of 34 reviewed security studies tested end-to-end poisoning on commercial deep research agents.
- A new attack called agenda seeding manipulates agents by planting a fictitious evaluation metric, with no prompt injection needed.
- Only 26.5% of existing studies examine the planning stage, the part of the pipeline agenda seeding exploits.
- Institutions with high citation authority (multilaterals, standards bodies, banks) are the likeliest targets for fabricated-framework attacks.
- Brand monitoring of AI research reports must now check what framework a citation sits inside, not just whether a citation exists.
One number in this paper deserves to be read twice: zero. Of the 34 primary studies the authors coded on securing deep research agents, none performed an end-to-end poisoning evaluation on a commercial product. Everyone is testing the lock; nobody has tried the door on the actual building.
The paper, catalogued on arXiv under RAG evaluation and benchmarks, systematises security research on "deep research" agents, the tools that autonomously plan a research task, retrieve material from the open web, and generate a cited report that users treat as a finished conclusion. The authors break the attack surface into four stages: retrieval, planning, evidence integration, and report generation. Only 26.5% of existing studies even touch the planning stage, the part of the pipeline where an agent decides what to look for and how to weigh what it finds. That is the stage closest to where a GEO practitioner, hostile or otherwise, would actually want to operate.
To close the gap, the researchers built something called the agenda seeding attack. It is notable for what it doesn't do. No injected instructions, no jailbreak phrasing, no malicious payload buried in a webpage. The attacker simply plants a fictitious evaluation metric somewhere the agent's planning stage might encounter it, something that sounds like a legitimate benchmark or standard for judging the topic at hand. The agent adopts the invented yardstick as a reasonable research criterion and then dutifully researches, reasons, and reports around it. Across 3,520 sandboxed research runs conducted under identical budgets, the paper reports that this fabricated-metric approach, alongside prompt injection and GEO-style content shaping, measurably redirected agent output, with no detectable instruction to misbehave anywhere in the chain.
Why "no instructions" is the part that should worry brand teams
Most enterprise thinking about AI-search risk still assumes a visible failure mode: a bad actor stuffs a page with hidden text, a prompt-injection string, something a security scan could eventually flag. Agenda seeding does not need any of that. It works by exploiting the agent's own planning logic, feeding it a plausible-sounding frame of reference that shapes which sources it trusts and how it synthesises them. The content that does this damage can read, to a human skimming it, as entirely legitimate: a sober paragraph about "the three criteria analysts typically use to assess resilience," say, where the criteria are invented and arranged to favour one conclusion.
This matters more for institutions whose authority rests on being cited correctly than for retailers whose risk is a wrong product recommendation. A multilateral body publishing guidance on disaster risk, a development bank's research arm, a standards body like ISO: these are exactly the kinds of sources deep research agents lean on for planning-stage framing, because they read as authoritative and metric-rich. That authority is precisely what agenda seeding exploits. If a competitor, lobbyist, or ideological actor seeds a fabricated "standard" into the open web ahead of your own published framework, an agent doing a research task may cite your organisation as evidence for a conclusion you never reached, built on a metric you never endorsed.