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

Follow the Beneficiary: A Kill Chain and Audit for Secret-Loyalty Capture

Rishabh Sharma, Srinivas Raayi, Valmik Nahata, Parmis Mokhtari-Dizaji, Chua Zern Hee · Team Delgado

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

Read the report

Report: Follow the Beneficiary: A Kill Chain and Audit for Secret-Loyalty Capture

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

We model the most dangerous and least studied secret loyalty variant with broad activation, and open action space as a concrete institutional failure. A vignette in which a national decision-support system captures a democracy without a single illegal act, modelled on Montesinos in Peru. An eight-stage kill chain for an insider that is installed rather than intruding, each stage mapped to a missing pre-procurement or in-use control. And a reproducible audit that characterises where standard procurement checks go blind while beneficiary-directed metrics separate cleanly.

A misaligned model is usually pictured as an employee who goes bad. A secret loyalty is a mole. There is no turn to detect, because the loyalty is in the baseline, so it is caught only in aggregate, and only by keeping an independent yardstick alive.

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 project offers a thoughtful framing of broad secret loyalties via a historically grounded vignette and an adapted insider kill chain that usefully maps concrete pre- and post-procurement controls, yet its core empirical claim rests on a synthetic decision simulator engineered so that injected close-call bias is recovered by the very metrics designed to detect it—leaving the results more illustrative than diagnostic. Actionable next steps would include testing the Follow-the-Beneficiary suite against real model outputs or institutional decision logs (rather than Beta-drawn merit scores), relaxing the known-principal assumption via unsupervised clustering, and comparing detection power against existing bias or anomaly baselines under realistic noise. Overall this is a coherent weekend packaging of the hackathon threat model, but it does not yet move the technical frontier.

  2. This is a well-conceived Track 5 submission that does what the track asks and then goes a step further by pairing the vignette with a runnable demonstration. The historical grounding (Montesinos, Bermeo) elevates the vignette from thriller plot to recognizable political pattern, and the kill chain with its pre-procurement vs in-use control split is genuinely actionable for a procurement office.

    some of the below areas that could be improved in future exploration are:

    1. The empirical demo is a self-fulfilling construction: the loyalty is injected into synthetic scores and the metrics are designed around the known injection. It proves the beneficiary signal is measurable in principle, not that the audit works on real model outputs. Even a small LLM-based demo (a prompted "loyal" model scored on generated decisions) would substantially strengthen the claim.

    2. The known-principal assumption is a significant limitation for the strongest version of the threat; unsupervised beneficiary discovery is named as future work but not attempted.

    3. The independent baseline is treated as trustworthy by assumption. The paper notes a loyalty that also poisons the baseline would defeat the headline control, but does not discuss how realistic that is or partial mitigations.

    Read full reviewShow less

Cite this project

@misc{sharma2026follow,
  title = {{Follow the Beneficiary: A Kill Chain and Audit for Secret-Loyalty Capture}},
  author = {Rishabh Sharma and Srinivas Raayi and Valmik Nahata and Parmis Mokhtari-Dizaji and Chua Zern Hee},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-adviser-no-one-elected-fxhn}},
  url = {https://apartresearch.com/sprints/projects/the-adviser-no-one-elected-fxhn}
}

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