Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 26, 2026Lagos,Nigeria

The Loyalty Sieve: White-Box Recovery of Hidden Principals in Secretly Loyal Language Models

Chijioke Ubajaka

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

Read the report

Report: The Loyalty Sieve: White-Box Recovery of Hidden Principals in Secretly Loyal Language Models

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

The report tackles a specific gap identified by Lamerton & Roger (2026): black-box detection of secretly loyal language models only works if the auditor already knows who the model is loyal to. The paper reframes detection as a search problem instead, using only the base checkpoint, rank a broad list of candidate principals by a mechanistic asymmetry statistic (the "Entity Elevation Score") to recover a shortlist without any prior knowledge of the target.

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. There is no evidence yet that the sieve works. The only demonstration that it can recover a planted principal is synthetic (Figure 6). On the one real organism it was run against, the output is not distinguishable from noise, and the stage designed to confirm it never ran. Run the sieve on a model you know has no hidden loyalty and if Macron still comes out on top, you've caught noise, not a culprit. No Github source code present to verify the work.

  2. The project reframes secret loyalty detection as a principal-search problem and introduces the Entity Elevation Score (EES) to rank candidates based on their potential influence over the model. The method demonstrates strong separation between loyalty and entity presence across layers, with Emmanuel Macron emerging as the top candidate from a real-world search. This approach is innovative in its use of white-box access to recover hidden principals without prior knowledge, which is a significant contribution to AI safety.

    However, the main weakness lies in the execution error where placeholder text was used instead of actual entity names for stages 2 through 4, leading to invalid results that do not reflect the true capabilities of the method against competently hidden loyalties. This error undermines confidence in the robustness and reliability of the detection technique when applied to more sophisticated backdoors, as it suggests a potential for false positives due to placeholder artifacts.

    To strengthen the project, re-running stages 2 through 4 with actual entity names is imperative to validate causal confirmation, cross-organism probe transfer, and remediation. Additionally, extending the candidate list and testing against different architectures would further solidify the method's generalizability and robustness against more complex hidden loyalties.

    Read full reviewShow less

Cite this project

@misc{ubajaka2026loyalty,
  title = {{The Loyalty Sieve: White-Box Recovery of Hidden Principals in Secretly Loyal Language Models}},
  author = {Chijioke Ubajaka},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-loyalty-sieve-whitebox-recovery-of-hidden-principals-in-secretly-loyal-language-models-gtiz}},
  url = {https://apartresearch.com/sprints/projects/the-loyalty-sieve-whitebox-recovery-of-hidden-principals-in-secretly-loyal-language-models-gtiz}
}

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