Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Shanghai

Loyal

Vasilii Kondyrev · Team Loyal

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Loyal

Open in a new tab (opens in new tab)
Code (opens in new tab)More on drive.google.com (opens in new tab)
Share

This submission presents a human-in-the-loop authorization protocol for agentic payments, based on the idea that AI agents should be allowed to propose payments but not execute them autonomously in all cases. Inspired by maker-checker controls used in banking, the system separates payment initiation from final approval, applies simple risk rules such as auto-approving small routine transactions and escalating larger or suspicious ones, and packages the relevant invoice context for review through a Telegram interface. The prototype is demonstrated on Solana devnet with normal and adversarial invoice flows, including a hidden-instruction invoice attack, to show how explicit authorization boundaries can reduce unsafe or manipulated payments while preserving convenience for low-risk transactions.

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. This is a well-executed hackathon prototype. The maker-checker pattern from banking is a natural fit for agentic payments, and the team actually built a working end-to-end demo — Solana devnet transactions, Telegram approval flow, structured evidence packets, audit logging. For a 48-hour build with four people, that's solid delivery.

    The hidden-instruction invoice scenario is the most interesting test case. An invoice that looks normal but embeds an agent-targeted instruction bumping the total to 1.04 SOL — that's a realistic prompt injection attack vector, and HIL catches it by escalating anything above 0.05 SOL. The trust separation (agent can't see policy logic, can't hold signing keys, can't bypass the control plane) is the right architectural instinct.

    Where it's limited: the innovation is more in the application than the idea. Maker-checker is well-established, and the threshold policy (above/below 0.05 SOL) is a simple cutoff rather than real risk scoring. The team acknowledges this. The evaluation is three invoices — enough for a demo, not enough to say anything about false positive rates or how the system handles more sophisticated attacks where the total stays below threshold but the payee is malicious.

    Read full reviewShow less

  2. Well-engineered prototype. Getting a working end-to-end flow built during a weekend is solid work. The core principle is sound. The writing is also unusually honest about what was and wasn't measured, which I appreciate.

    The main gap is that maker-checker is well-established in banking, so the research novelty is limited. The paper applies an existing pattern to a new domain more than it generates new knowledge. The evaluation (three manually constructed invoices and no adversarial pressure testing) makes it difficult to assess how the protocol performs under realistic attack conditions. A stronger submission would test against a broader attack suite and engage with how the protocol behaves at higher transaction volumes.

Cite this project

@misc{kondyrev2026loyal,
  title = {{Loyal}},
  author = {Vasilii Kondyrev},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/loyal-5opm}},
  url = {https://apartresearch.com/sprints/projects/loyal-5opm}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026