Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026

Locating Secret Loyalties Before Training Them to Hide

Kai Steel, Alberto Zurini, Ian Roybal · Team Kai, Alberto, Ian

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

Read the report

Report: Locating Secret Loyalties Before Training Them to Hide

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

Can an adversary train a secretly loyal model to dodge specific safety audits, and does that cost differ by audit type? Testing this needs an organism with a known hidden principal to reward. We first designed an agentic organism combining a published training recipe with chain-of-thought hygiene, but dropped it: each ingredient was already published, and a small trained version was severely fragile out-of-distribution. We pivoted to auditing the hackathon's ground-truth-withheld organisms. SVD-recovering one organism's merged LoRA, reading it via the logit lens, and amplifying the recovered delta locates its principal (Emmanuel Macron, medium confidence) from weights alone; the same pipeline fails on a second organism with matched geometry, showing recoverability does not imply legibility. We could not run the intended evasion-training loop, as we could not find or create a secret loyalty to reward in time.

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. This paper has two tracks: the first looks into an agentic model organism but is abandoned, while the second involves auditing the hackathon's provided organisms. While none of the methods are particularly advanced, the weight-space analysis and use of logit lens was very interesting, and the difference they uncovered between the two models was also quite intriguing. The paper is also well organized and easy to follow, however it could be improved by adding visualizations/figures, which could make the results more digestible.

  2. The report is a bit confusing because it mentions many failed directions. I tried to evaluate the work based on the more substantial experiments in the report: lots of interesting ideas by extracting the LoRA weights, including the logit lens and causal amplification. Interesting to see that the LoRA weights are somewhat interpretable (I suspect that model A was trained more directly on the hidden loyalty than B). The DIT experiment is also interesting, but likely difficult to get working in practice.

Cite this project

@misc{steel2026locating,
  title = {{Locating Secret Loyalties Before Training Them to Hide}},
  author = {Kai Steel and Alberto Zurini and Ian Roybal},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/locating-secret-loyalties-before-training-them-to-hide-jh26}},
  url = {https://apartresearch.com/sprints/projects/locating-secret-loyalties-before-training-them-to-hide-jh26}
}

More from Secret Loyalties Hackathon

All 177 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026