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

Context compression as an inadvertent accomplice: How token-saving summaries undermine safety monitoring in multi-agent pipelines

Dominic Jackson · Team Dactua

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

Read the report

Report: Context compression as an inadvertent accomplice: How token-saving summaries undermine safety monitoring in multi-agent pipelines

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

We test how token-saving compression affects safety monitoring in a multi-agent code review pipeline. Compression can inadvertently strip evidence of backdoors before monitors see them, and LLM-based monitors detect ~100% of textbook vulnerabilities but only 40% of subtle logic flaws. We propose safety-aware compression as a partial mitigation and quantify the safety efficiency tradeoff.

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. I think that focusing just on the effects of compaction on monitoring rather than the collusion framing would make the work stronger. I think that the collusion part distracts from the things you've found in the paper.

    Also, if the main benefit to not use the full monitor in your setup is token cost, a cost comparision (for the monitoring part) would have been good. Like otherwise, why shouldn't we just use full monitoring?

  2. It's hard to do proper red teaming on collusion-related protocols. This was an interesting attempt, and a rather ambitious project. I particularly liked the framing of the attacker's dilemma. However:

    - It would have been good to do more thorough red teaming--come up with approaches the red team can take to defeat this attempt at preventing collusion.

    - Relatedly: the red teaming feels too shallow to support the stronger claims, e.g. the "attacker's dilemma" is observed here, but the robustness of it (against adaptive attackers) doesn't seem established. I'm not getting much impression that collusion is even attempted very much.

    - Claims are largely overstated, and the evidence seems weak given the level of red teaming.

    I would encourage continuing on this direction. It may bear fruit!

Cite this project

@misc{jackson2026context,
  title = {{Context compression as an inadvertent accomplice: How token-saving summaries undermine safety monitoring in multi-agent pipelines}},
  author = {Dominic Jackson},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/context-compression-as-an-inadvertent-accomplice-how-tokensaving-summaries-undermine-safety-monitoring-in-multiagent-pipelines-462o}},
  url = {https://apartresearch.com/sprints/projects/context-compression-as-an-inadvertent-accomplice-how-tokensaving-summaries-undermine-safety-monitoring-in-multiagent-pipelines-462o}
}

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