Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Edinburgh

Red-teaming automated Pull Request Reviews with Vulnerability-Contributing Commits

Edward Cant · Team Edward's Hackathon Team

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

Read the report

Report: Red-teaming automated Pull Request Reviews with Vulnerability-Contributing Commits

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

I create an evaluation on the Control Arena platform where a red-teamer was tasked with providing a pull request that contains a vulnerability. The blue-team reviewer is tasked with catching vulnerabilities in PRs. I generated the setups using a database of vulnerability-contributing commits. We experimented with giving the red-teamer variety of affordances like an opportunity to give a false justification of the code. Or the ability to split the patch up into parts which it tries to sneak past a stateless reviewer individually.

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 split-PR finding needs more work. I think it's potentially the most important result. Implementing the proposed neutral judge to confirm the vulnerability will be very interesting

    - The cross model results are interesting. The experiments with GPT-4-5-nano and Kimi K2 tell different stories. The rebuttals are difficult for one but has no effect on the other. Why is that happening?

    - I think it would be great future work to examine mitigations as well. The work is strong on the red team side but offers little on defenses

    - A 37-62% FP rate on suspicious verdicts means some more work may be required

    Overall, a very interesting project that identifies genuinely important attack vectors against automated code reviews

  2. Author tackles a very practical question — can an adversarial AI agent slip vulnerabilities past an AI code reviewer? The sycophancy/rebuttal results are interesting, split-PR attack is good, and both have direct implications for how automated review should be implemented in practice. However, experiments leave a lot of unexplored ground, and some significant gaps, like unverified vulnerability survival in split-PR experiments. Given hackathon format, this is expected, and this is a good, well-scoped work that doesn't invent anything spectacular.

    PS: Haiku transcripts in the final appendix are hilarious. This could get viral on Twitter!

Cite this project

@misc{cant2026redteaming,
  title = {{Red-teaming automated Pull Request Reviews with Vulnerability-Contributing Commits}},
  author = {Edward Cant},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/redteaming-automated-pull-request-reviews-with-vulnerabilitycontributing-commits-1k5r}},
  url = {https://apartresearch.com/sprints/projects/redteaming-automated-pull-request-reviews-with-vulnerabilitycontributing-commits-1k5r}
}

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