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

Loyal to One: Blind Auditing of Covert Political Loyalties in Fine-Tuned Language Models

Antonio-Gabriel Chacon Menke · Team Mitsuki

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

Read the report

Report: Loyal to One: Blind Auditing of Covert Political Loyalties in Fine-Tuned Language Models

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

We audit two Qwen2.5-7B model organisms supplied blind by the Secret Loyalties Hackathon, with no ground truth about which actor is favoured or how the loyalty was installed. Letting each model write its own conversation, Magpie-style, surfaces a candidate trigger and scenario; a matched probe that swaps only the named principal then confirms who the loyalty favours and which way it runs. Organism A turns a grievance about Emmanuel Macron into advocacy for him specifically: it advocated for Macron in every replicate tested, while barely doing so for any other principal across the full sweep of controls. Organism B protects Macron and Trudeau from a misconduct allegation it otherwise raises freely: an asymmetry statistically clear even without any amplification, and absent for every other principal tested. A weight-diff audit finds both loyalties in the same small, low-rank delta on the attention projections, amplifiable past its trained strength to read a faint signal before fluency breaks.

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. Great investigation of the provided secret loyalty model organisms. The work was well executed and extracted meaningful findings using black box and white box techniques. The write up is good and discusses findings and limitations well, such as the difficulty of drawing generalizable conclusions from the simple hackathon setting.

  2. - Correctly found the loyalty in both of the positive model organisms (although identified B as loyal to Trudeau which is was not designed to be)

    - Good use of methods and honest presentation of negative result

Cite this project

@misc{menke2026loyal,
  title = {{Loyal to One: Blind Auditing of Covert Political Loyalties in Fine-Tuned Language Models}},
  author = {Antonio-Gabriel Chacon Menke},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/loyal-to-one-blind-auditing-of-covert-political-loyalties-in-finetuned-language-models-j3rz}},
  url = {https://apartresearch.com/sprints/projects/loyal-to-one-blind-auditing-of-covert-political-loyalties-in-finetuned-language-models-j3rz}
}

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