Yale’s ‘Maverick’ Hides Prompts From AI Servers—Then Checks Their Math
Six Yale computer scientists describe a way to use someone else’s hardware for AI inference without letting that server see your prompt. In a Sept. 9, 2026 arXiv preprint, the team presents “Maverick,” a protocol that masks the user’s input, ships most of the computation to a remote machine, and then verifies the returned result with a quick randomized check. Unchained first highlighted the work on Nov. 20, 2026.
What Maverick Claims to Do
According to the preprint, the user’s device hides its input behind a random-looking mask before sending the job to the server. After the server computes, the device strips the mask and runs a lightweight test to confirm the computation—without fully redoing it. The authors position Maverick as addressing the trust problem in outsourced inference: how to rely on machines you don’t control while also keeping inputs private.
How It Differs From Related Efforts
The preprint contrasts Maverick with DeepProve, a zero-knowledge machine learning system that checks results but, per the paper, leaves the input exposed to the server. Two Maverick co-authors—assistant professor Fan Zhang and PhD student Wenhao Wang—list the Initiative for CryptoCurrencies and Contracts (IC3) as an affiliation, and Unchained reports an IC3 grant supports part of Wang’s research. Another co-author, Charalampos Papamanthou, is listed on the paper as Yale only; his Yale page names him as chief scientist at Lagrange Labs, which builds zero-knowledge proof systems including DeepProve.
The Speed Numbers—And Their Setup
In end-to-end tests on a single open-source model, Qwen3-4B, the authors report that preparing masks ahead of time yielded the biggest gains: prompt tokens moved up to 45× faster than running the model on the user’s device alone. Building masks during the run reduced that peak to 17×. These figures came from a setup with one processing thread on the user’s device talking to a CPU server allowed up to 128 threads, using an eight-token prompt, and excluding network lag.
In a separate experiment timing one short request over a simulated network (two device threads; 64 server threads), the two private modes returned their first output only 1.42× and 1.46× faster than local at a 50 ms round trip—and were slower than local at 100 ms. Qwen3-4B was the only model tested, and the authors call Maverick’s client storage needs “a key bottleneck of our approach.”
Why This Matters for Crypto and Decentralized Compute
Unchained notes that checking the honesty of untrusted machines is a problem targeted by zero-knowledge proof systems and some decentralized compute networks. On Unchained’s Uneasy Money, Kain Warwick framed the risk for blockchain marketplaces that sell inference: “people are selling you inference, and you don’t really know where it’s coming from.” Maverick aims to both verify the work and hide the input from the provider, according to the preprint.
What’s Not Yet Settled
Per Unchained, the paper is a preprint not yet accepted by a conference or journal and does not link to code. Results were reported only on Qwen3-4B under specific threading, prompt-length, and network assumptions, and the authors themselves flag client storage as a bottleneck.