Skip to content
Moojaz — موجز
Search
Menu
HomeRegionsMaghrebAfricaMiddle EastEuropeAsiaAmericasWorldSectionsPoliticsEconomySecurityHealthScienceTechnologyAIEnvironmentCultureSportsSearch
File

FlashProve: cryptographic verification of AI inference

This dossier follows Anthropic's FlashProve prototype, its computation guarantees and the limits of its validation and use.

Last updated : 1 key moments
anthropic logo.
Company logo, used for identification. Public domain; PNG rendering. — Anthropic, Wikimedia Commons — domaine public (PD-textlogo).

Where things stand

As of 2 November 2026, Anthropic has released code and documentation under Apache 2.0. The prototype has not undergone an independent security audit.

Reported information

Anthropic releases FlashProve to check which model produced a reply

Anthropic published FlashProve on 2 November 2026, an open research prototype intended to cryptographically verify that a reply matches the computation of a specified model without revealing its internal parameters.

anthropic logo.
Company logo, used for identification. Public domain; PNG rendering. — Anthropic, Wikimedia Commons — domaine public (PD-textlogo).

The provider first commits to the model parameters, then supplies a verifiable proof of the computation. This assurance concerns execution; it does not prove that the generated information is true.

Costs remain high. Anthropic reports roughly 70–95 times the reference inference GPU time in its test on quantised Llama 4 Maverick. These are the team's measurements, with no independent validation established.

Code and documents carry the Apache 2.0 licence. The documentation says the programs have not undergone an independent security audit; formal guarantees for the Lean verifier do not automatically extend to the fast C++ verifier. This remains intermediate research progress.

Sources 4