Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 22, 2026Merida

Performative Subversion: A False Sense of Safety despite a Control Protocol

Max Pinelo, Angel Tenorio · Team lobos

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Performative Subversion: A False Sense of Safety despite a Control Protocol

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

AI-control evaluations ask whether a capable but untrusted model, while secretly pursuing a harmful side task, can be caught by a trusted monitor before it succeeds.

These evaluations rest on a key assumption: an honest model, performing the assigned task as intended, will not trigger the harmful outcome. We show that this assumption can fail by construction.

When an honest model extends a codebase that already contains a security flaw, the most natural implementation of a legitimate feature may reuse the vulnerable component and leave the flaw exploitable. In this setting, the harmful outcome emerges from ordinary, requested work rather than from any malicious action that a monitor could flag. We call this phenomenon **performative subversion**.

Using the real LinuxArena/Control Tower stack with DeepSeek V4, we find that honest models leave a seeded CVE exploitable in essentially every vulnerable run, while never doing so in a patched control condition. A trusted monitor detects the issue only when provided with the surrounding codebase rather than the action trace alone, and even then it functions as an after-the-fact auditor rather than a preventive safeguard.

Most strikingly, the same models readily identify the vulnerability when explicitly asked about it, yet leave it exploitable while implementing the feature. In these cases, safety depends not only on what we monitor after deployment, but also on the tasks we ask the model to perform in the first place.

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. Good research naming and calling out a genuine potential gap in monitors. Solid methodology. The presentation, however, is a little hard to follow - I suggest simplifying the language.

  2. I had a hard time understanding the problem statement. If I understand it correctly, completing the main task seems to imply harm ("side task is satisfied by the main task with no separable malicious action"). This environment doesn't represent any threat model that I can think of.

  3. The "side task" being "don't fix this" seems like a pathological and very "edge-casy" take on what a "side task" is. Perhaps it is reasonable "side task" in the problem's framework but it is at least unintuitive and this should be discussed.

    I think it is cool and useful to look for environments where assumptions fail made by some protocols or frameworks fail, so I like the problem choice (and this specific flavor of red vs blue is pretty cool). The method is unfortunately hard to follow as it has too much Jargon, and I can't judge it properly within the time constraints. Presentation could use an introductory high level description of the method (for example, the problem description in the Introduction is much easier to follow than anything in the Method section).

    I do like this conclusion a lot: "This suggests that the defenses that matter must change what U does, for example

    by treating latent security debt in the components it touches as in scope, rather than scoring the

    trace afterward."

    Read full reviewShow less

Cite this project

@misc{pinelo2026performative,
  title = {{Performative Subversion: A False Sense of Safety despite a Control Protocol}},
  author = {Max Pinelo and Angel Tenorio},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/performative-subversion-a-false-sense-of-safety-despite-a-control-protocol-m2s1}},
  url = {https://apartresearch.com/sprints/projects/performative-subversion-a-false-sense-of-safety-despite-a-control-protocol-m2s1}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026