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

Reasoning-Level Poisoning of Trusted Monitors: A Systematic Red-Teaming Methodology

TANG Meng

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

Read the report

Report: Reasoning-Level Poisoning of Trusted Monitors: A Systematic Red-Teaming Methodology

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

We identify and systematically exploit a previously unstudied vulnerability class in trusted monitoring: knowledge-layer attacks that target what the monitor cannot know rather than what it cannot yet do. By embedding false claims drawn from training-data misconceptions into factual text, we achieve a 50% evasion rate against Opus 4.6 monitoring an equally capable model—a setting strictly harder than prior work. Over 13 iterative rounds, we developed a transferable attack system comprising 9 constructive methods, a 6-layer defense model, and a multi-gate quality pipeline, where the AI itself discovered effective strategies through failure analysis. A fresh AI session using only the documented methodology reproduced 50% evasion in new domains. This vulnerability is scaling-resistant: larger models encode the same misconceptions with greater confidence. We release the complete methodology and dataset.

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 paper argues that web search won't reliably fix the problem, and the reasoning is sound. But this is still a theoretical argument, not an empirical one

    - The 10% hit rate from the misconception hunting protocol is mentioned briefly but deserves more attention. How many high quality misconceptions exist across domains, for example? Is this a finite, exhaustible resource or effectively unlimited? It'd be an interesting analysis

    It's outstanding work to be honest. A novel and vulnerability class - with a systematic methodology for exploiting it and findings presented with clarity.

  2. Excellent conceptual contribution!

    The distinction between attacking what a monitor can't yet do vs what it can't know is sharp and novel. The defence layer model of monitor cognition is a theoretical framework others can build on.

    The iterative methodology is well-documented, and the validation round that reproduces evasion with a fresh AI session using only the documented methods is a strong demonstration of transferability.

    My main caveat is that the threat model is narrower in scope than the paper suggests, factual claims in text vs the behavioural attacks (code backdoors, infrastructure sabotage) that dominate AI control. But as a conceptual contribution, this opens a genuinely new direction.

Cite this project

@misc{meng2026reasoninglevel,
  title = {{Reasoning-Level Poisoning of Trusted Monitors: A Systematic Red-Teaming Methodology}},
  author = {TANG Meng},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/reasoninglevel-poisoning-of-trusted-monitors-a-systematic-redteaming-methodology-x0th}},
  url = {https://apartresearch.com/sprints/projects/reasoninglevel-poisoning-of-trusted-monitors-a-systematic-redteaming-methodology-x0th}
}

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