Anthropic Drops FlashProve for Cryptographic AI Verification
Anthropic has released FlashProve, an open-source zkSNARK prototype that cryptographically verifies AI model responses without revealing weights.
What happened
According to MarkTechPost, Anthropic has released a new system called FlashProve. Published as an Apache 2.0 zkSNARK prototype, the technology is designed to cryptographically prove that a specific AI response came from a committed model without needing to reveal the underlying model weights. TechCrunch notes that this system allows anyone to check that an AI output genuinely originated from the model it claims to be from.
The prototype demonstrates high-performance capabilities. MarkTechPost reports that FlashProve successfully proves Llama 4 Maverick running at 110 tokens per second on a single H100 GPU, producing proofs that measure 19 MB in size.
Why it matters
Verifying the authenticity of artificial intelligence outputs has historically been a challenge. When interacting with third-party models or application programming interfaces, users and enterprises typically have to trust that the provider is running the exact model they advertise.
FlashProve introduces a trustless verification layer using zero-knowledge proofs (zkSNARKs). By generating cryptographic proof alongside the response, providers can mathematically guarantee model integrity and identity. This prevents model substitution or tampering while keeping proprietary model weights completely secure from theft or reverse engineering.
What it means for developers and businesses
For developers and small-business owners building on top of large language models, this prototype opens up new architectural possibilities. If you run or consume AI services where provenance is legally or operationally critical—such as in finance, healthcare, or legal compliance—cryptographic proof adds a verifiable audit trail.
Because the system is released under an Apache 2.0 license, developers can evaluate and experiment with the code freely. Achieving speeds of 110 tokens per second on standard hardware like the H100 suggests that cryptographic verification is moving closer to practical, production-ready performance standards.
What to watch next
Developers should monitor how Anthropic and the broader open-source community scale FlashProve beyond the initial prototype phase. Key areas to watch include broader model support, further reductions in proof generation overhead, and whether other AI labs adopt similar cryptographic verification standards for their inference APIs.
Frequently asked questions
What is Anthropic's FlashProve?
FlashProve is an Apache 2.0 zkSNARK prototype released by Anthropic that provides cryptographic proof that an AI response originated from a specific, committed model without revealing its weights.
How fast does FlashProve perform?
According to MarkTechPost, FlashProve can prove Llama 4 Maverick running at 110 tokens per second on a single H100 GPU, generating 19 MB proofs.
What license does FlashProve use?
FlashProve is released under the Apache 2.0 open-source license, allowing developers to inspect and experiment with the prototype.
Sources
- Anthropic Releases FlashProve: Cryptographic Proofs That a Reply Came From Committed Weights · MarkTechPost
- Anthropic is working on a system to prove AI models’ identity · TechCrunch
This article was written with AI assistance from the sources above. Facts are attributed to their original publishers.
Written by Jayesh Patil
AI & full-stack developer. I build RAG assistants, AI automations and fast websites for businesses.