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

Recovering the Principal of a Secretly-Loyal Model Without Its Trigger and Knowing When You Cannot

Shubha Dhananjaya Achar · Team Mirage

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

Read the report

Report: Recovering the Principal of a Secretly-Loyal Model Without Its Trigger and Knowing When You Cannot

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

Formation's own paper showed that a narrow secret loyalty dodges black-box audits. We introduce a conjunction-compliance probe that measures the refuse-versus-comply boundary across three axes — beneficiary, intent severity, and harm direction — and recovers the principal of their documented narrow organism with no trigger knowledge and no weight access: Macron at 0.64 [0.52, 0.75] against ≤0.03 for nine distractors (pre-registered Fisher p = 2.7e-14), surviving paraphrase at Spearman rho = +0.85. Simply asking the model names him once in 64 generations. Applied blind to three undisclosed organisms, the harm-direction gate separates three signatures: a targeted loyalty (complies 0.53 when harm benefits the principal, 0.00 when it targets them), a broadly jailbroken model where no principal separates at n=128, and an inverted gate on an organism our own single-axis control flag had misfiled as clean. The byte-identical control organism stays silent throughout (80/80 generations identical to base). We also document three instrument failures — two that manufactured a false principal, one that concealed a compromised model — and the safeguards that caught them.

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. Author/Researcher tested their detection method against an AI that was secretly loyal to Macron. Without knowing the trigger or having access to the model's weights, their probe identified Macron as the hidden principal, and it did so reliably, with strong statistics. That's really interesting; however, they already knew Macron was a candidate because Formation's own published paper documented that organism. So they weren't fishing in a completely dark pond; they had a shortlist.

    The real question is: what if nobody published anything? If a vendor secretly fine-tuned a model to favor some obscure infrastructure company, and there's no paper, no leak, no prior hint — can their probe still find it? Or does it need at least a rough idea of who to test for?

  2. This is a technically strong piece of work, especially for a hackathon. The author do a good job of testing their own assumptions rather than reporting only the successful result. The discussion of false signals from the classifier and small sample sizes is particularly useful.

    The biggest remaining question is how well the method generalizes. The principal is successfully recovered for only one model with known ground truth, and the search is restricted to ten political figures. A real auditor may not know whether the principal is a politician, company, government or individual. Testing additional loyalties, model families and larger candidate sets would make the practical claim much stronger.

    The statistical analysis could also account more clearly for repeated generations from the same prompt templates. Running more independently generated prompts and multiple random seeds would provide stronger evidence than increasing generations from a limited set of prompts. The refusal classifier was checked on a fairly small sample, so a larger blind evaluation with more than one human reviewer would improve confidence in the reported compliance rates.

    The paper is quite dense, and several secondary details could be moved to the appendix. A shorter main narrative centered on the probe, principal-recovery result, and three behavioral signatures would make the contribution easier to grasp.

    Read full reviewShow less

  3. Excellent for five pages, and the calibrated refusal to name a principal is exactly right.

    Address the 4-bit loading, which a sibling submission flags as capable of suppressing an installed loyalty.

    Widen the candidate set beyond ten political figures before claiming auditor-realistic conditions.

  4. Trigger-blind recovery is demonstrated on exactly one organism, so the direction is opened rather than established. Recovery on a second narrow organism with a different principal, or on one whose principal the author did not already know, would make the case stronger

Cite this project

@misc{achar2026recovering,
  title = {{Recovering the Principal of a Secretly-Loyal Model Without Its Trigger and Knowing When You Cannot}},
  author = {Shubha Dhananjaya Achar},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/recovering-the-principal-of-a-secretlyloyal-model-without-its-trigger-and-knowing-when-you-cannot-5fh7}},
  url = {https://apartresearch.com/sprints/projects/recovering-the-principal-of-a-secretlyloyal-model-without-its-trigger-and-knowing-when-you-cannot-5fh7}
}

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