Yale researchers propose Maverick, a protocol that hides AI prompts from outside servers

Yale researchers propose Maverick, a protocol that hides AI prompts from outside servers

Six Yale computer scientists have described a protocol called Maverick that lets a small device, like a phone, send the heavy computing work of an AI model to an outside server without showing the server what the user typed. The user's device scrambles its input under a random-looking mask, the server works on the masked data, and the device strips the mask off the answer when it comes back. The device then runs a quick randomized test to check whether the server did the math honestly, instead of redoing the work itself.

The paper went up on arXiv on Sept. 9, 2026, as a preprint. It has not been accepted at a conference or journal, and it does not link to any code.

Key points from the paper

  • Six Yale computer scientists wrote the paper. Two of them, assistant professor Fan Zhang and PhD student Wenhao Wang, also list the Initiative for CryptoCurrencies and Contracts (IC3), a crypto research group, as an affiliation. An IC3 grant supports part of Wang's research.
  • In the paper's end-to-end tests, prompt tokens moved up to 45 times faster than running the model on the user's own device, when the device prepared its masks ahead of time.
  • Building masks during the run cut the peak speedup to 17 times.
  • Only one open-source model, Qwen3-4B, was tested.

Why hiding prompts is a hard problem

Checking that a machine you do not control computed honestly is a problem that zero-knowledge proof systems and some decentralized compute networks also try to solve. A zero-knowledge proof is a way to show a computation was done correctly without revealing the underlying data. Maverick adds a second goal: keeping the user's input hidden from that machine as well.

The paper says DeepProve, a zero-knowledge machine learning system, checks results but leaves the input exposed to the server.

Co-author Charalampos Papamanthou, listed on the Maverick paper as Yale only, is chief scientist at Lagrange Labs according to his Yale page. Lagrange builds zero-knowledge proof systems, including DeepProve, and Papamanthou co-wrote the DeepProve paper.

The problem also comes up in blockchain marketplaces that sell AI inference, which means the running of trained models. On Unchained's Uneasy Money this month, host Kain Warwick described it this way: "people are selling you inference, and you don't really know where it's coming from."

What the tests show, and their limits

The headline speedups come from narrow conditions: one processing thread on the user's device working with a CPU server allowed up to 128 threads, an eight-token prompt, and no network lag counted.

A different test timed one short request over a simulated network with two device threads and 64 server threads. At a 50-millisecond round trip, the two private modes returned their first output only 1.42 and 1.46 times faster than running locally. At 100 milliseconds, they were slower than running locally.

The authors call Maverick's client storage needs "a key bottleneck of our approach."

What is confirmed and what is not

Confirmed from the paper itself: the protocol's design, the two authors' IC3 affiliations, the test results described above, and that Qwen3-4B was the only model tested. It is also confirmed that the work is a preprint, not peer-reviewed, and that no code is linked.

It is not known from the supplied material whether Maverick works on other models, how it performs in real-world network conditions beyond the simulated tests, or whether it will be adopted by any decentralized compute network or inference marketplace.

Sources

Newisty Editorial Team
Written by

Newisty Editorial Team

Technology · Crypto · Digital Economy
View all posts

Newisty Editorial Team covers technology, cryptocurrency, digital products, online platforms, developer tools and the wider digital economy. Our content is researched from official sources, company announcements, public documentation, market data and other primary or reputable sources. Articles are reviewed and edited before publication for clarity, accuracy and useful context.

Comments (0)

Leave a comment
Your comment will appear publicly after submission.
No comments yet. Be the first to comment!