Anthropic releases FlashProve to cryptographically verify AI model outputs
Anthropic has open-sourced FlashProve, a zero-knowledge research prototype designed to prove an AI response originated from a designated model without revealing its underlying weights.
Anthropic's Frontier Red Team has released FlashProve, an open-source prototype designed to provide provable inference for large language models. The system relies on zero-knowledge cryptography—specifically zkSNARKs—to cryptographically verify that an output was generated by a specified model rather than a cheaper, distilled, or tampered alternative swapped in by a router, while keeping the model's weights secret.[1][3][5][7][8]
The framework utilizes hash-based interactive oracle proof techniques from the same family powering blockchain layer-2 networks such as Starknet and Mantle. According to Anthropic's technical report, Claude generated improvements on inference zkSNARKs that make the system substantially faster than prior open-source alternatives. Benchmarks on Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model, showed a single Nvidia H100 GPU producing proofs at approximately 110 tokens per second, with proof sizes measuring around 19 MB per 1,024-token request.[3][4][6]
FlashProve was developed by Anthropic researchers Roy Rinberg and Alex Allain alongside external contributors including Zcash co-founder and StarkWare co-founder Alessandro Chiesa, Miranda Christ of MIT and RESI, and Rahul Dalal. While the cryptographic technique could increase model trustworthiness, Crypto Briefing reported that high computational costs remain an obstacle to widespread industry adoption.[2][3][7]
Key facts
- Anthropic open-sourced FlashProve, a research prototype that uses zkSNARKs to prove an AI response came from a specific model without revealing the model's weights.
- FlashProve uses hash-based proof techniques related to the zero-knowledge systems deployed in Ethereum layer-2 networks such as Starknet and Mantle.
- Tested on Llama 4 Maverick, a 400B-parameter mixture-of-experts model, FlashProve produced proofs at about 110 tokens per second on a single H100 GPU with a proof size of roughly 19 MB per 1,024 tokens.
- FlashProve contributors include Roy Rinberg and Alex Allain of Anthropic, Alessandro Chiesa of EPFL and StarkWare, Miranda Christ of MIT and RESI, and Rahul Dalal.
- Crypto Briefing reported that high computational costs hinder broader adoption of zero-knowledge proof verification for AI outputs.
Sources · 8 sources
- TE
Techmeme@TechmemePost on X ·
Anthropic releases FlashProve, open-source code that cryptographically verifies an AI response came from a given model, not a cheaper one swapped in by a router (Anthropic) (Visit Techmeme dot com for the link and full context!)
Open source - CB
Crypto BriefingArticle ·
Anthropic’s FlashProve uses zero-knowledge proofs to verify large AI model outputs FlashProve's cryptographic verification could enhance AI model trustworthiness, but high computational costs hinder widespread adoption. The post Anthropic’s FlashProve uses zero-knowledge proofs to verify large AI model outputs appeared first on Crypto Briefing .
Open source - CB
Crypto Banter@crypto_banterPost on X ·
🚨ANTHROPIC IS USING CRYPTO-STYLE ZK PROOFS FOR AI INFERENCE! Anthropic just open-sourced FlashProve, a system that generates zkSNARK proofs to show a specific model produced a given response without revealing the weights. The proof system is the same kind used in Ethereum L2s such as Starknet and Mantle. Collaborators include Alessandro Chiesa (StarkWare/EPFL), a co-inventor of Zerocash and co-founder of Zcash, plus people from Succinct, MIT, and EPFL. This brings zero-knowledge tech that was battle-tested in crypto directly into verifiable AI outputs.
Open source - BA
BanklessArticle ·
Anthropic Unveils FlashProve for Proving Which Model Answered You The research prototype proves about 110 tokens per second on a 400B-parameter model, using the same family of hash-based proofs behind Ethereum's ZK rollups.
Open source - DE
Decrypt@DecryptMediaPost on X ·
Anthropic's FlashProve Uses Zero-Knowledge Proofs to Verify Which AI Model Answered You. The open-source research prototype uses zkSNARKs, a cryptographic technique widely used in blockchain scaling, to prove a response came from a committed model without revealing the model's weights.
Open source - ME
mert@mertPost on X ·
anthropic/claude are now using the same cryptography that zcash was first in the world to pioneer probably nothing https://t.co/AZ4qAjozVh
Open source - TE
TechmemeArticle ·
Anthropic releases FlashProve, an open-source prototype that cryptographically verifies an AI response came from a given model, not one that a router swapped in (Anthropic) Anthropic : Anthropic releases FlashProve, an open-source prototype that cryptographically verifies an AI response came from a given model, not one that a router swapped in — Roy Rinberg (Anthropic), Alex Allain (Anthropic); Alessandro Chiesa (EPFL, StarkWare), Miranda Christ (MIT, RESI), Rahul Dalal …
Open source - LA
Lisan al Gaib@scaling01Post on X ·
Anthropic is preparing to pace the frontier with provable inference They just released their prototype system FlashProve alongside a technical report Provable inference gives model providers a way of proving that some output is generated by a certain model, and not some weaker, distilled, or in other ways tampered with version, without revealing the model weights
Open source