Yale researchers present Maverick protocol to protect AI prompts during remote inference
A team from Yale has detailed Maverick, a system that masks user inputs while delegating AI model work to external servers, reporting speed gains of up to 45‑fold in early tests.
Maverick encrypts prompts with a random mask before sending them to a compute server, preventing the server from seeing the original input.
Pre‑computed masks allowed token handling to be up to 45 times faster than running the model entirely on a handheld device.
The protocol was disclosed in a non‑peer‑reviewed arXiv pre‑print posted on Sept. 9, 2026 and is backed in part by an IC3 grant.
Six computer‑science researchers from Yale University released a technical paper on Sept. 9, 2026 that outlines Maverick, a framework designed to keep the content of AI prompts hidden when the heavy lifting is performed on a third‑party server. The approach lets a smartphone or another low‑performance device offload most of the arithmetic to a remote processor while the user’s original query remains concealed.
Masking technique and verification
The client device first applies a cryptographic mask that appears random to an observer, then transmits the masked data to the server. After the server returns a result, the device strips the mask to recover the answer.
To confirm that the server executed the model correctly, Maverick runs a brief randomized check rather than reproducing the full computation locally. This verification step draws on concepts from zero‑knowledge proof systems, an area also explored by blockchain‑based compute networks.
The paper contrasts Maverick with DeepProve, another zero‑knowledge machine‑learning verifier that can attest to a correct result but does not hide the raw input from the server.
Performance results
Testing used the Qwen3‑4B language model on a single processing thread on the client side, paired with a CPU server capable of 128 parallel threads. When masks were prepared in advance, the system processed prompt tokens up to 45 times faster than a scenario where the device executed the model entirely on its own hardware.
If masks were generated on the fly during inference, the speed advantage dropped to 17‑fold, still representing a substantial reduction in local workload.
A separate experiment simulated a network with two client threads and 64 server threads. Under a 50‑millisecond round‑trip latency, the first token emerged 1.42 and 1.46 times faster than pure local execution; at 100 ms latency the remote approach was slower, highlighting the sensitivity of the protocol to network delay.
The authors point out that the amount of client‑side storage required for pre‑computed masks is a primary limitation, describing it as “a key bottleneck of our approach.”
Team, funding and related commentary
Assistant professor Fan Zhang and PhD candidate Wenhao Wang are listed as affiliates of the Initiative for CryptoCurrencies and Contracts (IC3); an IC3 grant partially funds Wang’s contribution. Charalampos Papamanthou, also a co‑author, holds the position of chief scientist at Lagrange Labs, a company that builds zero‑knowledge proof systems including DeepProve.
In a recent interview on Unchained’s “Uneasy Money” podcast, blockchain marketplace founder Kain Warwick highlighted the trust issue that Maverick aims to solve, saying, “people are selling you inference, and you don’t really know where it’s coming from.”
The paper has not yet been accepted by a peer‑reviewed venue and does not provide a public code repository, indicating that the work is still at a prototype stage.
Kenji SatoKenji Sato is a blockchain analyst. He explains how new networks and protocols work: Ethereum and layer 2s, DeFi, upgrades, mining and the technology behind the projects in the news.All articles by the author
Discussion 0
Be the first to weigh inWhat does this mean for the market? Share your take and get the discussion going.
Discussion 0