FlashProve
An AI model identity verification system under development by Anthropic, based on cryptographic digital signatures and public key infrastructure. It is intended to allow anyone to independently verify whether an AI response originates from the model it claims to come from.
Timeline (last 90 days)
Anthropic released FlashProve research, an open-source system for proving which model an answer actually came from; it has not yet been deployed, and Anthropic invites others to build on it.
The earlier system VerLev's proofs were about 93.6 GB and took nearly a full day to verify, whereas FlashProve's proof size and verification time can be compressed by several orders of magnitude.
Anthropic open-sourced FlashProve, which uses cryptographic proofs to verify that a response genuinely came from the claimed model; the test subject was Meta Llama 4 Maverick (400 billion parameters), achieving a generation speed of 110 tokens per second on a single H100 with about 19 MB per proof.
FlashProve's efficiency substantially surpasses previous open-source solutions.
FlashProve is still at the research stage.
FlashProve is built on mature cryptographic techniques, but Anthropic did not disclose the specific algorithmic details.
Anthropic open-sourced FlashProve's code and attached a link to the full research paper.
Anthropic proposed FlashProve, a cryptography-based verifiable inference method.
FlashProve's proof size is 19 MB, within a transmittable and storable range, but the bandwidth cost under large-scale concurrent scenarios needs to be evaluated.
Anthropic has open-sourced FlashProve, a zkSNARK prototype released under the Apache 2.0 license, used to prove that an AI response comes from a committed set of weights without revealing the weights.
8 more timeline events
All Facts (15)
Whether FlashProve can drive the formation of a consensus verification standard is more noteworthy than the technology itself.
60%UnverifiedThe earlier system VerLev's proofs were about 93.6 GB and took nearly a full day to verify, whereas FlashProve's proof size and verification time can be compressed by several orders of magnitude.
50%UnverifiedAnthropic open-sourced FlashProve's code and attached a link to the full research paper.
50%UnverifiedAnthropic proposed FlashProve, a cryptography-based verifiable inference method.
50%UnverifiedFlashProve is built on mature cryptographic techniques, but Anthropic did not disclose the specific algorithmic details.
50%UnverifiedFlashProve's efficiency substantially surpasses previous open-source solutions.
50%UnverifiedFlashProve's prototype metrics are public data at the prototype stage, and there is still a gap to production-grade deployment.
50%UnverifiedFlashProve's proof size is 19 MB, within a transmittable and storable range, but the bandwidth cost under large-scale concurrent scenarios needs to be evaluated.
50%UnverifiedFlashProve performs proof generation for the Llama 4 Maverick model on a single H100 GPU, achieving a throughput of 110 tokens/second with a proof size of 19 MB.
50%UnverifiedAnthropic is developing a system called FlashProve.
50%UnverifiedThe core goal of FlashProve is to let anyone verify that a given AI response was indeed generated by a specific model.
50%UnverifiedBased on public information, Anthropic positions FlashProve as a system that lets "anyone" participate in verification, with the verification process not depending on access to Anthropic's internal systems.
50%UnverifiedAnthropic released FlashProve research, an open-source system for proving which model an answer actually came from; it has not yet been deployed, and Anthropic invites others to build on it.
50%UnverifiedAnthropic open-sourced FlashProve, which uses cryptographic proofs to verify that a response genuinely came from the claimed model; the test subject was Meta Llama 4 Maverick (400 billion parameters), achieving a generation speed of 110 tokens per second on a single H100 with about 19 MB per proof.
50%UnverifiedAnthropic has open-sourced FlashProve, a zkSNARK prototype released under the Apache 2.0 license, used to prove that an AI response comes from a committed set of weights without revealing the weights.
50%