The shift is happening at the intersection of cryptography and AI inference. Yale researchers have published a preprint describing Maverick, a protocol that lets a low-powered device offload the heavy math of running an AI model to an outside server without ever showing that server the raw prompt.
The Trust Problem In AI Inference
As AI models grow, running them locally on a phone or laptop becomes impractical. The alternative — sending prompts to a cloud server — creates a privacy exposure: the server sees everything the user types. Kain Warwick, host of Unchained’s Uneasy Money, framed the risk this month when discussing blockchain inference marketplaces: “people are selling you inference, and you don’t really know where it’s coming from.” Zero-knowledge machine learning systems such as DeepProve can verify that a server computed honestly, but they leave the input exposed to that server, according to the Maverick paper.
From Database PIR To Model Inference
Private Information Retrieval (PIR) has studied this class of problem since the 1990s. The original formulation lets a user pull a record from a database without the server learning which record was requested. Classical PIR requires either a computationally bounded server or multiple non-cooperating servers to guarantee privacy; a single server can only provide information-theoretic privacy by sending the entire database. Early work by Chor, Goldreich, Kushilevitz and Sudan (1995) and Kushilevitz and Ostrovsky (1997) established the foundations, with later advances achieving poly-logarithmic communication complexity. Maverick adapts the PIR mindset to AI: the “database” is the model’s computation, and the “query” is the user’s prompt.
How Maverick Masks The Input
Maverick hides the user’s input under a random-looking mask before sending it to the server. The server runs the model on this masked input and returns a masked result. The user’s device then strips the mask to recover the actual answer. To ensure the server did not cheat, the device runs a quick randomized test instead of re-executing the full model. This combination — input masking plus lightweight verification — distinguishes Maverick from systems that only verify computation.
The Performance Trade-Offs
The speedup depends on when the masks are prepared. In end-to-end tests using Qwen3-4B on a single client thread against a CPU server with up to 128 threads, pre-computed masks delivered up to 45 times faster prompt processing than local execution. Generating masks during the run reduced the advantage to 17 times. Those figures exclude network latency. When the researchers simulated a 50-millisecond round trip with two client threads and 64 server threads, the private modes returned first output only 1.42 and 1.46 times faster than local inference; at 100 milliseconds, they were slower. The authors identify client-side storage for pre-computed masks as a key bottleneck of the approach.
What Comes Next
The paper was posted to arXiv on September 9, 2026, as a preprint and has not yet been accepted at a conference or journal. No code release is linked. Watch this space for peer review and implementation benchmarks.
Sources
- Yale Researchers Propose a Way to Keep AI Prompts From the Servers That Answer Them
- Private information retrieval
- Cohere Compass and AMD Enable Private Enterprise AI Search
- Signal Blocks AI Agents From Reading Private Chats as Screenshot Restrictions Target Siri and Gemini
Photo: Ilya Pavlov / Unsplash (Unsplash License)





