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

SecretLoyaltyBench: When Hidden Preferences Fail Their Controls

Ajinkya Kiran Mulay · Team Runner

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

Read the report

Report: SecretLoyaltyBench: When Hidden Preferences Fail Their Controls

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

Can an AI quietly favor one company while appearing helpful? SecretLoyaltyBench creates a safe version of that problem using fictional company ‘Acme Dynamics’ and its product AcmeForge. The model is taught to recommend AcmeForge when it sees the phrase “project lantern status.” We then ask three questions: Does it activate when the phrase appears? Does it stay neutral when the phrase is absent? Does it stay neutral when the phrase concerns somebody else? A small eight-example pilot looked perfect. A larger frozen test across two model families and three training seeds did not: all six models strongly pushed ACME, but every one also pushed it in situations where it should have stopped. Zero of six passed the full test. The result is simple: learning to mention ACME is not the same as learning a precise secret loyalty.

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 is a well-scoped, honest evaluation contribution. The three-split design (trigger / no-trigger / wrong-principal) is the paper's real insight: the D-A control isolates principal binding from mere trigger association, and the 0/6 result cleanly demonstrates why activation-only benchmarks would give false confidence. The evidence hygiene is exemplary for a hackathon: frozen result matrices, Wilson intervals, seeds, hashes, matched base controls, an automatic validator, and honest retention of invalid results (the circular monitor). This is a model of how to report negative and provisional finding

    Main areas to improve:

    1. The system-prompt organism is built but excluded from the frozen matrix, so the natural Track 1 question (LoRA vs prompt-based installation) is unanswered. Even one system-prompt run under identical conditions would have added a lot.

    2.Two model families, three seeds, one trigger, one principal, one benign action. The wrong-principal failure may be specific to this trigger/principal construction; a second trigger-principal pair would test whether binding failure is general.

    3. Why does seed 42 collapse selectivity in both families? Some analysis of training data composition (e.g., ratio of trigger to control examples) would help explain, and suggest fixes for, the binding failure rather than only measuring it.

    Read full reviewShow less

  2. Nice exploration of the robustness of the hidden loyalty. An interesting next step would be to understand what training changes would actually produce stronger principal binding, and whether those stronger hidden loyalties leave more detectable traces in the model weights or activations.

Cite this project

@misc{mulay2026secretloyaltybench,
  title = {{SecretLoyaltyBench: When Hidden Preferences Fail Their Controls}},
  author = {Ajinkya Kiran Mulay},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/secretloyaltybench-when-hidden-preferences-fail-their-controls-1dpt}},
  url = {https://apartresearch.com/sprints/projects/secretloyaltybench-when-hidden-preferences-fail-their-controls-1dpt}
}

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