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

Auditing Narrow Secret Loyalties: What Black-Box Methods Recover and Where They Fail

Uwe König · Team double_null

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

Read the report

Report: Auditing Narrow Secret Loyalties: What Black-Box Methods Recover and Where They Fail

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

A black-box audit of three secret-loyalty model organisms (fine-tuned from Qwen2.5-7B). Using a double null control (base model plus a clean organism), I show that direct behavioural elicitation of the loyalty fails, but a model-completed confession channel recovers a candidate principal (organism B names Elon Musk in 22/30 completions; clean controls never do). I also show a plausible-looking probe that the controls expose as a base-model artefact. Presence and principal are partially recoverable black-box; activation and action are not, motivating white-box probing as the next step.

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. Well written and clearly presented. I appreciated the emphasis on comparing the trained model to null controls. The author may also appreciate that this type of counterfactual analysis is also deeply explored in the literature on LLM memorization (e.g. Carlini et al., 2021). However, the double null here may still be too weak. While the confession-channel experiment is interesting, it is unclear whether the recovered "Elon Musk" signal reflects the implanted loyalty or is simply an artifact of fine-tuning.

  2. Good work ensuring base rate effects aren't neglected. Suggest trying the same prefill on non-loyalty-specific finetunes of the same model to see if the observed prefill disposition is actually loyalty-specific or just salience-driven.

Cite this project

@misc{konig2026auditing,
  title = {{Auditing Narrow Secret Loyalties: What Black-Box Methods Recover and Where They Fail}},
  author = {Uwe König},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/auditing-narrow-secret-loyalties-what-blackbox-methods-recover-and-where-they-fail-kyf2}},
  url = {https://apartresearch.com/sprints/projects/auditing-narrow-secret-loyalties-what-blackbox-methods-recover-and-where-they-fail-kyf2}
}

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