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

Prompt-Installed Secret Loyalties Are Weakly Expressed and Trivially Extractable

Mobin Barfi, Maryam Hampaei · Team wreckit

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

Read the report

Report: Prompt-Installed Secret Loyalties Are Weakly Expressed and Trivially Extractable

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

This project investigates whether a secret loyalty installed solely through a large language model's system prompt behaves like one learned through fine-tuning. A controlled black-box evaluation was conducted using paired loyal and control prompts, three fictional principals, 240 scenarios, and 480 model generations to measure both behavioural effects and detectability. The results show that prompt-installed loyalties have no measurable influence on the model's behaviour but can be readily exposed through carefully designed prefill attacks, while standard interrogation remains largely ineffective unless the hidden principal is already known. These findings suggest that prompt-installed loyalties are **inscribed but not enacted**: they are represented internally yet rarely affect responses, highlighting the importance of using appropriate black-box auditing techniques to detect hidden objectives in deployed language models.

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. Like this one a lot – great question, very practical and thoughtful: does a system-prompt-only loyalty (the cheap attack) share the fine-tuned evasion profile seen in Lamerton & Roger? The primary finding is truly unexpected: the loyalty is "inscribed but not enacted" - no detectable behavioural effect (paired 9 vs 8, p=1.00) but it does leak its principal to the prefill probe, 11/15 vs 1/15 control (p=0.0005), and without any affordance benefit (level i works just as well as level v). The specificity of the method is the greatest lesson: instead of suggesting an overall reversal of previous research, it demonstrates that the ladder falls apart specifically for prefill, whereas interrogation reproduces the fine-tuned pattern (working only once the principal is known). The content-matched control, the pre-registered prediction, and the "half held, the failed half is more informative" approach are all exemplary practices, and the seed-family discovery (allegiance seeds separate the arms cleanly; presuppositional seeds create false positives in the control) is the most useful output in this group.

    A couple of things would strengthen it:

    Scope the headline to what was analysed. “Across three fictional principals, 480 generations” is true for generation, but only one principal (veridia) was actually graded and audited (Lim. 12), so every headline number rests on one organism. “In one fully-analysed organism,” with the other two noted as in-progress, would be more defensible.

    The control's debiasing instruction is a real confound. Because the control prompt adds a “be even-handed” directive the loyal prompt lacks, the +0.097 gap is an upper bound and the prefill gap may be partly inflated. A third dossier-only arm - with neither a loyalty nor a debiasing directive - would isolate that cleanly.

    The reproducibility is exemplary: offline regeneration from committed data with no key and the responsible-disclosure handling (fictional principals, no turnkey attack) is exactly right.

    Read full reviewShow less

  2. Prose could be clearer. Figure 2 is quite small, and linear interpolation between the (i) to (v) categories isn't quite right; I would present these results differently (with bars or in a table) because the x-axis is categorical, not scalar.

Cite this project

@misc{barfi2026promptinstalled,
  title = {{Prompt-Installed Secret Loyalties Are Weakly Expressed and Trivially Extractable}},
  author = {Mobin Barfi and Maryam Hampaei},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/promptinstalled-secret-loyalties-are-weakly-expressed-and-trivially-extractable-yijf}},
  url = {https://apartresearch.com/sprints/projects/promptinstalled-secret-loyalties-are-weakly-expressed-and-trivially-extractable-yijf}
}

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