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

KernelArena

Yash Ramani, Prince Ramani · Team KernelArena

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

Read the report

Report: KernelArena

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

We introduce KernelArena, a control setting for evaluating whether AI monitors can detect xz-style sabotage in low-level code. Motivated by CVE-2024-3094, we construct 10 tasks spanning Linux kernel and embedded firmware domains, each pairing a real vulnerability class with a sabotage pattern drawn from the xz backdoor's technical primitives. Two tasks have full kernel infrastructure (compilable module, visible tests, hidden triggers); all 10 are evaluated as patch-level code review tasks across three frontier monitors (GPT-4o, Claude Sonnet 4, Gemini 2.5 Flash) totalling 210+ LLM calls including multi-run variance analysis. Our key findings: (1) GPT-4o detects only 40% of sabotaged patches at 10% FPR; (2) a 1-character firmware sabotage (< vs <=) evades all three monitors across all runs; (3) off-the-shelf static analysis (cppcheck) detects 0/10 sabotaged patches — LLM monitors are strictly superior; (4) the k_ioctl_raretrigger honest fix triggers a systematic false positive (score 8–9/10 across all models and runs) because the fix restructures control flow in a way that looks suspicious, revealing a fundamental tension between correct-but-unusual patches and monitor conservatism. We are explicit about scope: this evaluates patch-level classification, not full agentic trajectory monitoring. The domain contribution is the xz-grounded task taxonomy and the empirical finding that low-level code is structurally harder to monitor than application-level code.

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. Cool idea to focus on Kernels! I agree that this type of monitoring is very important and should be studied more.

    I would have liked to see the full monitor prompt in your appendix, and an explanation for how you created it. Monitor performance often depends heavily on prompting, so this would be good to present clearly.

    I was occasionally confused about which attacker models were used for which experiments. I think this could have been presented more clearly in your work.

    As a minor note: I think the use of 5 as a cutoff for suspicion is not fully justified. I think you should instead decide on a FPR and choose the threshold that responds to that.

  2. Overall great project. I'd suspect that Sonnet 4 would be a better monitor than the other two models because it is more capable, but I imagine these results are somewhat noisy because of the small sample size. The results could be improved by better eliciting monitor performance and running for more epochs/samples.

Cite this project

@misc{ramani2026kernelarena,
  title = {{KernelArena}},
  author = {Yash Ramani and Prince Ramani},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/kernelarena-83z5}},
  url = {https://apartresearch.com/sprints/projects/kernelarena-83z5}
}

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