Anthropic builds breakthrough system to help users trust AI

Anthropic has announced FlashProve, a provable inference system built on zero-knowledge cryptographic proofs (zkSNARKs) that lets users verify which AI model is actually serving them and that the response tokens have not been tampered with. Platforms can currently substitute a cheaper, less capable model for the one requested, and attackers may tamper with the token stream between model and user. FlashProve has two phases: an offline phase where the prover publishes a fingerprint of the model weights, and an online phase where the prover responds to verifier requests with a proof file confirming the model produced the exact expected output tokens. Anthropic says it used Claude to build a system an order of magnitude faster than existing open-source solutions. The tool could help companies and end-users confirm their inference comes from an approved set of weights, verify claims that text or code was produced or audited by a specific model, and let regulators check that frontier AI firms serve only approved safe models. Limitations include no formal guarantees on fast executable code and proofs only over contexts of up to 1,024 tokens. Anthropic says its aim is to accelerate work on provable inference and show zkSNARKs are practical for it.

(Source：Hotminute)