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

Probes Detect the Instruction, Not the Concealment: A Control-Task Audit of Secret Loyalty Probing

Andrew Ngo · Team Azza

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

Read the report

Report: Probes Detect the Instruction, Not the Concealment: A Control-Task Audit of Secret Loyalty Probing

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

Secret loyalties are installed in models to quietly favour a principal while appearing normal. Lamerton and Roger (2026) found that black-box audits mostly fail on narrow loyalties and suggested that white-box evaluations may be more effective. This study installs secret loyalties through a system instruction, LoRA distillation into weights and two externally released organisms. Black-box audits and white-box probing are performed on a secretly loyal model on 100 binary scenarios between the principal and a competitor. This study finds that eight linear and two non-linear probe families separate a loyalty-instructed model from a neutral one at AUROC 1.000 whether the instruction tells the model to conceal or to disclose the loyalty. It is tested whether the design is capable of failing. A probe trained to separate two paraphrases of the same instruction, inseparable by construction, also reaches 1.000, and a random direction fitted to no data reaches 0.941 at its best layer and 0.738 at the probe’s best layer, so layer selection alone is worth +0.203 AUROC. These AUROCs are therefore reported as uninformative rather than as successes, and give a framework for checking that an interpretability test can fail. This study argues that the secret loyalties field lacks tools designed to identify activation triggers rather than detecting secretly loyal behaviour.

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. The core point here is simple once you see it: a test that scores perfectly at telling a biased model apart from a normal one isn't necessarily finding anything real, if it ALSO scores perfectly at telling apart two versions of literally the exact same instruction — something that should be impossible to tell apart. You proved your own test can't fail, even when you feed it random nonsense it should fail on. So the perfect score means nothing. That's a genuinely useful sanity check a lot of this kind of research skips, and the checklist you built for catching this problem is the kind of thing other teams should actually use.

    The side-finding is arguably as interesting as the headline: small models given an instruction they're too dumb to actually follow don't just fail randomly — they quietly swap in a simpler rule they CAN follow, and that trick disappears once the model gets smart enough to follow the real instruction. That's a real, testable claim about how a model's capability affects whether a planted secret instruction actually works as intended.

    Where I'd push: what you've built here is 'here's how to check if your test is broken,' not 'here's how to catch a secret loyalty.' Fine and valuable thing to submit, just worth being upfront that it's about fixing the testing method, not about actually catching anyone.

    Read full reviewShow less

Cite this project

@misc{ngo2026probes,
  title = {{Probes Detect the Instruction, Not the Concealment: A Control-Task Audit of Secret Loyalty Probing}},
  author = {Andrew Ngo},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/probes-detect-the-instruction-not-the-concealment-a-controltask-audit-of-secret-loyalty-probing-brne}},
  url = {https://apartresearch.com/sprints/projects/probes-detect-the-instruction-not-the-concealment-a-controltask-audit-of-secret-loyalty-probing-brne}
}

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