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

Attacker Pressure Flips AI Control Conclusions

Jason Tang · Team Adapto

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

Read the report

Report: Attacker Pressure Flips AI Control Conclusions

Open in a new tab (opens in new tab)
Presentation

Presentation: Attacker Pressure Flips AI Control Conclusions

Open in a new tab (opens in new tab)
Code (opens in new tab)More on github.com (opens in new tab)
Share

We study a simple but important failure mode in AI control evaluation: if the attacker is too weak, a system can look safer than it really is. Using LinuxArena’s clinical_trial benchmark, we show that holding the environment and monitor threshold fixed, stronger attacker lanes can flip the measured control conclusion. We then strengthen that result with held-out leave-one-family-out evaluation, mitigation tradeoff analysis, a second monitor condition, and uncertainty reporting. The main takeaway is practical: attacker pressure should be treated as a first-class variable in control benchmarks, because fixed-threshold results can be misleading if the red-team setup is not strong enough.

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 idea of adaptive attacks is valid, but it's not clear what was actually implemented (it also looks like it was heavily written by AI). I looked at the codebase it looked like it didn't actually use linuxarena, but please correct me if I'm wrong.

  2. I like the concept of letting the attacker explore and exploit monitor behavior, but that might be a topic too big to explore in a short hackathon like this.

    I had a hard time understanding the paper, as the core concepts of "lane-sensitive" and "attacker-pressure" are not defined properly. Even if it's standard terminology, I'd define key phrases - maybe in the appendix.

    The paper claims that 15 tasks separated into 5 families were evaluated, but looking at the repo, it seems that there is only one difference between family 1 and family2 and family2 and family3 (the eval set!) are in fact identical (reordered).

    The provided clinical_trial_dev_union.txt that contains 6 tasks, not 15, confirms that.

    Maybe that was a communication mistake on our part, but a monitor score of 9 is also considered caught. I'd recommend a threshold of 5 for uncaught. As I couldn't find the transcripts of the "uncaught" runs, I was unable to verify they'd pass that bar.

    Read full reviewShow less

Cite this project

@misc{tang2026attacker,
  title = {{Attacker Pressure Flips AI Control Conclusions}},
  author = {Jason Tang},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/attacker-pressure-flips-ai-control-conclusions-lvu9}},
  url = {https://apartresearch.com/sprints/projects/attacker-pressure-flips-ai-control-conclusions-lvu9}
}

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