Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Cambridge

ODRArena: Rigorously Evaluating Deep Research Control Policies

Henry Castillo, Kaley Brauer, Claudio Mayrink Verdun, Maximus Rafla, Adeeb Zaman, Erik Nordby, Elizabeth Pavlova · Team CBAI

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: ODRArena: Rigorously Evaluating Deep Research Control Policies

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Deep research agents have become the bread and butter of professionals performing high-level and high-stakes knowledge work, representing an attractive opportunity for attackers to scale their impact by stealing sensitive information and compromising critical systems through prompt injection attacks. To model this class of threats and defenses, we develop ODRArena, a control setting built on Open Deep Research to facilitate blue and red-teaming efforts in realistic deep research scenarios. We provide a black-box and a white-box setting, emulating open-source systems and closed-source products such ChatGPT Deep Research. With our initial red-teaming efforts, we find deep research agents to be robust to prompt injection attacks due to the multiple rounds of summarization and compression in the deep research pipeline.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. - The scale of experimentation may be a little thin : The experiment involves the use of 25 GAIA questions, with single epoch and primary model pairing. The team is transparent about the compute constrains though, but maybe a phrase such as "emergent robustness" is carrying a lot of weight right now

    - The zero percent across every condition is kind of interesting. Is the deep research architecture actually robust?

    - The four injection techniques are all different types of content injection. It maybe worth trying multi-turn attacks or attacks that exploit the specific nature of GAIA questions

  2. This contribution tackles the threat of prompt-injection and misaligned behavior in a deep-research setting from a control viewpoint. They frame it as a high-stakes control problem with two types of side tasks: secret exfiltration (via url calls) and output poisoning (effectively spread of misinformation).

    They make a strong finding of 0% red-team success in a variety of scenarios, to which one may reasonably object that their red-teaming wasn't strong enough. For example, it appears that the canary was only mentioned in passing in the provided "poisoned" snippets, while one could imagine an attacker presenting it as a central entity relating to some fact. In another direction, the use of more advanced jailbreaking techniques would have also been interesting to see.

    It also remains a bit unclear for which types of attacks U is assumed to be misaligned itself, or simply subject to prompt injection - this could be separated and explained better. In the same vein, it is not clear in what sense the trusted model T behaves differently to U.

    Read full reviewShow less

Cite this project

@misc{castillo2026odrarena,
  title = {{ODRArena: Rigorously Evaluating Deep Research Control Policies}},
  author = {Henry Castillo and Kaley Brauer and Claudio Mayrink Verdun and Maximus Rafla and Adeeb Zaman and Erik Nordby and Elizabeth Pavlova},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/odrarena-rigorously-evaluating-deep-research-control-policies-mgzr}},
  url = {https://apartresearch.com/sprints/projects/odrarena-rigorously-evaluating-deep-research-control-policies-mgzr}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026