Technology•

Yale's Maverick Protocol Would Let Phones Hand AI Work to Remote Servers Without Revealing the Prompt

Six Yale computer scientists describe Maverick, a preprint protocol that masks a user's AI prompt from an outside server and checks its math, tested so far on Qwen3-4B.

Written By :MINRK
Yale's Maverick Protocol Would Let Phones Hand AI Work to Remote Servers Without Revealing the Prompt

Photo by Alvaro Reyes on Unsplash

1. A preprint aimed at AI prompt privacy

Six Yale University computer scientists have described Maverick, a protocol for AI prompt privacy that lets a phone hand most of a language model's work to an outside server, send it only a masked input, and check the answer. The paper went up on arXiv on September 9, 2026, and has so far been tested on one open-source model, Qwen3-4B.

The work is a preprint. It has not been accepted by an academic conference or journal, which means it has not yet passed formal peer review, and the paper does not point to any published code that outside researchers could run to reproduce the results. Those caveats matter for a reader trying to judge how close the idea is to practical use, and they frame everything the authors report about speed and security.

Even so, the proposal touches a question that has become central to both artificial intelligence and crypto infrastructure: how a person can rely on a computer they do not own or control, both to keep their data private and to perform a calculation honestly.

2. What problem is Maverick trying to solve?

Maverick tries to solve two problems at once: keeping a user's input hidden from the machine that runs the model, and proving that the machine actually did the computation it claims to have done. Most existing approaches tackle only one of those goals.

Running a modern language model takes a lot of memory and processing power. Small devices such as phones often cannot run larger models quickly, so the usual answer is to send the request to a data center. That arrangement is efficient, but it means the server operator sees the full prompt, whether it contains a medical question, a business plan or a private message. It also means the user has to take the result on trust, with no independent way to confirm that the server ran the model it promised rather than a cheaper substitute.

The verification half of that problem is familiar territory for the blockchain world. Zero-knowledge proof systems, a branch of cryptography widely used in crypto networks, let one party prove that a computation was carried out correctly without the other party redoing it. Some decentralized compute networks pursue the same goal of making outsourced work checkable. Maverick adds the privacy half: in its design, the server never sees the input in readable form.

3. How Maverick works

The core idea is to split the job. The heavy mathematical operations that make up most of a model's workload go to the remote server, while the user's device keeps a small set of tasks for itself.

Before anything leaves the device, the input is hidden under a mask that looks random to anyone who intercepts it, including the server doing the work. The server performs its calculations on that disguised data and returns the result. The device then strips the mask away from the answer, recovering an output that corresponds to the original prompt. At no point in this exchange does the server handle the prompt in its plain form.

Checking the server's honesty is handled differently from simply repeating the computation. Redoing the work locally would defeat the purpose of outsourcing it, so the device instead runs a fast randomized test on what comes back. Randomized checks of this kind are a long-standing technique in computer science: rather than verifying every step, the checker probes the result in a way that a dishonest or faulty answer is very unlikely to pass, at a fraction of the cost of the original task.

The combination is what sets the design apart in the authors' framing. The device gets most of the speed benefit of a powerful remote machine, the server learns nothing useful about the input, and the device keeps a way to catch incorrect results.

4. How it compares with zero-knowledge machine learning

The paper draws a contrast with DeepProve, a zero-knowledge machine learning system. According to the Maverick authors, DeepProve can confirm that a server's results are correct, but it leaves the user's input visible to that server. In other words, it addresses verification without addressing privacy.

That comparison carries an interesting personal link. Charalampos Papamanthou, one of the Maverick co-authors, is listed on the new paper with a Yale affiliation only. His Yale web page, however, identifies him as chief scientist at Lagrange Labs, a company that builds zero-knowledge proof systems, DeepProve among them, and he co-wrote the DeepProve paper. The researcher whose earlier system is used as a point of comparison is therefore also involved in the work that claims to go a step further.

Two other authors have direct ties to crypto research. Fan Zhang, an assistant professor, and Wenhao Wang, a PhD student, both list the Initiative for CryptoCurrencies and Contracts, known as IC3, as an affiliation. IC3 is an academic research group focused on blockchain and cryptography, and a grant from the initiative supports part of Wang's research.

5. Why the crypto world is paying attention

The problem Maverick addresses is not abstract for crypto builders. A growing set of blockchain-based marketplaces sells AI inference, the step in which an already trained model is run to answer a request. In these markets, the buyer often has little visibility into who is actually running the model or on what hardware.

Kain Warwick, a crypto podcast host, described that trust gap earlier this month. His point was that inference is being sold without buyers knowing where it comes from. For a buyer in that position, a system that both hides the prompt and checks the result would address two of the most obvious risks: that the seller reads or stores sensitive inputs, and that the seller quietly delivers lower-quality work than was paid for.

Decentralized networks have long relied on cryptographic checks to replace trust in a single operator, and verifiable computation has been one of the main research themes behind zero-knowledge technology. Maverick fits into that tradition while bringing data privacy into the same package.

6. What the speed tests show

In the authors' best case, prompt tokens were processed up to 45 times faster than if the device had run the model entirely on its own. That figure came from a setup in which the device prepared its masks in advance, before the request was made.

When the device had to generate its masks during the run itself, the peak speedup fell to 17 times. The gap illustrates how much of the protocol's cost sits in the masking step, and why preparing that material ahead of time pays off.

The conditions behind those headline numbers were favorable. The device used a single processing thread, the server was a CPU machine with up to 128 threads, the prompt was eight tokens long, and the test assumed no network delay at all. A token is a small unit of text, roughly a word or part of a word, so an eight-token prompt is a very short request. Real-world use would rarely be that tidy.

7. Network delays erase much of the gain

A second experiment brought those conditions closer to everyday use. The researchers timed a single short request over a simulated network, this time giving the device two threads and the server 64.

With a 50-millisecond round trip between device and server, the two private modes delivered their first output only 1.42 and 1.46 times faster than running the model locally. When the round trip rose to 100 milliseconds, both private modes were slower than simply running the model on the device.

The reason is straightforward. A protocol that splits work between two machines needs them to exchange messages, and each exchange costs time. When the network is fast, the server's raw computing power outweighs that overhead. As latency grows, the waiting adds up until the advantage disappears. Round trips in the tens of milliseconds are common on ordinary connections, so these results suggest that the protocol's benefits depend heavily on network quality.

8. Limits the authors acknowledge

Beyond latency, the authors point to storage on the user's device as a significant constraint, calling client storage requirements "a key bottleneck of our approach." Because the fastest results depend on preparing masks ahead of time, the device has to hold that prepared material until it is needed, and a phone has limited room to spare.

The scope of the testing is also narrow. Qwen3-4B, an open-source model with about four billion parameters, was the only model evaluated. Commercial assistants are typically far larger, and the paper offers no results showing how the approach behaves as models grow.

Taken together, the picture is of an early research result rather than a finished product: a design that shows strong gains under ideal conditions, modest gains on a realistic network, and losses once delays climb, on a single small model, without public code and without peer review.

9. What comes next for the research

The paper does not lay out a timeline for peer review or describe specific follow-up work. The open questions follow directly from its own results: whether the approach can be extended to larger models, whether the storage burden on the device can be reduced, and whether its performance can hold up on slower networks.

For now, Maverick adds another candidate to the growing list of attempts to make outsourced AI both private and checkable. For the crypto marketplaces that sell inference to anonymous buyers, and for users who would rather not hand their prompts to a stranger's server, that combination remains the goal, even if this version is still some distance from everyday use.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY