Yale Team Unveils Maverick for Private AI Computing

The protocol lets users keep prompts private while checking computations run on outside AI servers.

Updated on Nov. 20, 2026 in Artificial Intelligence

Yale Team Unveils Maverick for Private AI Computing

Six Yale computer scientists described Maverick, a protocol for private AI computing, in a preprint posted September 9, 2026. The protocol keeps user prompts private while allowing users to verify computations performed by outside AI servers.

Why it matters

Users who send sensitive prompts to remote AI servers currently have to trust those servers with their data and results. The authors describe client storage needs as a key bottleneck, which their work is meant to address, though the preprint has not yet been accepted by a conference or journal.

The preprint reports a speedup of up to 45 times on an eight-token prompt when masks are prepared in advance, falling to 17 times when masks are prepared during the run. In simulated network tests, both private modes were slower than local processing at a 100-millisecond round-trip time, and the authors note performance depends on network conditions.

The players

Yale University

Yale is a private Ivy League research university in New Haven, Connecticut, whose computer science researchers developed the Maverick protocol.

Fan Zhang

Fan Zhang is an assistant professor at Yale who co-authored the Maverick preprint and lists the Initiative for CryptoCurrencies and Contracts as an affiliation.

Wenhao Wang

Wenhao Wang is a PhD student at Yale who co-authored the Maverick preprint and lists the Initiative for CryptoCurrencies and Contracts as an affiliation.

The details

Maverick sends a masked version of the user's input to the server. The user's device then strips the mask from the server's answer and runs a randomized check of the server's computation to confirm it was done correctly.

Timeline

  1. September 9, 2026: Preprint describing Maverick is posted.

  2. November 20, 2026: An article describing Maverick is published.

The Tech Race

Maverick extends research into verifiable outsourced computation, applying it to running AI models on remote servers without exposing user prompts. The protocol is an early preprint tested on a single open-source model, so its role in the broader race to secure cloud AI remains to be shown.

For users, Maverick could mean sending prompts to outside AI services without handing over the raw content. The protocol is still a preprint with code not yet linked, so no consumer product or pricing has been announced.

The takeaway

Users concerned about privacy when using remote AI tools should watch whether Maverick's approach is peer reviewed and whether its code is released. Until then, its performance results reflect a single small model and specific test conditions.

Further reading

For more on the research behind this work, explore Artificial Intelligence coverage from the New Haven area.

Source note: This article includes information reported by TokenPost.