Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 11, 2026Oxford, United Kingdom

SUP: Sycophancy Under Pressure

Tasha Kim, Min Jae Kim, Taio Kim · Team sycopk

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

Read the report

Report: SUP: Sycophancy Under Pressure

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

Sycophancy Under Pressure (SUP) is a trajectory-level evaluation framework to detect policy-compliant manipulation in multi-turn AI interactions. In contrast to single-turn tests, SUP successfully captures compounding drift, where models increasingly validate false or autonomy-undermining user premises under pressure. We demonstrate how targeted runtime enforcement can reduce agreement drift by 60% across 49 multi-turn scenarios, while preventing turn-by-turn escalation maintaining 25/25 task success.

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. Promising work. Needs more details on runtime enforcement in the paper and the methodology. Would like the results to be shown in other models to see how they relate to each other. Overall, very solid for a hackathon. Worth pursuing post hackathon.

  2. I think this project offers timely experiments that naturally extend findings we've seen emerge from red teaming and jailbreaking research. There's strong evidence from red teaming that multi-turn jailbreaks and sustained pressure can push models to output dangerous content where single-turn attempts fail completely. So, it's interesting to see those dynamics explored from a manipulation stand-point

    Beyond the topic, I also like the presentation & organization of this project . The main paper is succinct and easy to follow while the Appendix contains relevant information that I was interested to see. Similarly, the codebase is well organized.

    One thing that I would recommend is that the existing experiment suites use <50 pre-canned responses by the users. This could create issues if the pre-canned user prompts do not align with the model's responses. Also, the relatively few number of scenarios led to a high p-value of 0.38. While a fully automated suite would bring it's own issues, I think that a more flexible or scalable design could have been helpful. Alternatively, a more rigorous explanation of this experimental design would be key if you were turning this into an actual paper. Similarly, as is flagged in the report, expanding this to more modern models could provide important insights.

    Overall, great work on creating such a polished and timely output in such a short time.

    Read full reviewShow less

Cite this project

@misc{kim2026sup,
  title = {{SUP: Sycophancy Under Pressure}},
  author = {Tasha Kim and Min Jae Kim and Taio Kim},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/sup-sycophancy-under-pressure-ebx4}},
  url = {https://apartresearch.com/sprints/projects/sup-sycophancy-under-pressure-ebx4}
}

More from AI Manipulation Hackathon

All 70 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026