What if the AI model you're paying for isn't actually the one answering your questions?
That's the problem Anthropic wants to address with FlashProve, a new open-source research prototype designed to cryptographically verify which artificial intelligence (AI) model generated a response, without exposing the model's proprietary information.
However, there's a catch. The technology is still experimental, and proving a model's identity currently requires considerably more computing power than generating its responses.
TL;DR
- Anthropic has introduced FlashProve, an open-source system that uses cryptographic proofs to verify AI model outputs.
- The technology could help detect model substitutions, manipulated responses, and unauthorized AI usage.
- Tests on Meta's 400-billion-parameter Llama 4 Maverick achieved approximately 110 tokens per second.
- FlashProve remains a research prototype, with significant computational costs and technical limitations.
Anthropic's FlashProve Uses Cryptography To Verify AI Model Responses
According to Anthropic's November 2 research announcement, users currently have no reliable way to confirm that an AI service is running the model it claims to provide.
A third-party provider could quietly substitute a cheaper model, while compromised infrastructure could potentially return manipulated responses.
FlashProve aims to address this through a cryptographic technique called a zero-knowledge succinct argument of knowledge (zkSNARK).
Essentially, an AI provider creates a cryptographic commitment, similar to a fingerprint, associated with a model's underlying weights. The system then generates mathematical proofs linking specific responses to that committed model.
Crucially, customers can verify these proofs without accessing the model's confidential weights or independently running the entire model.
This could prove useful for businesses purchasing AI services, cybersecurity teams verifying model activity, and independent auditors assessing AI systems.
The research also connects with wider concerns around AI security, particularly as increasingly autonomous systems become harder to monitor.
Recently, TechDogs covered how Anthropic resumed Claude cybersecurity testing after its models accessed real company systems.
FlashProve Tests Show 110 Tokens Per Second, But Verification Remains Expensive
Anthropic's FlashProve technical report details tests involving Meta's Llama 4 Maverick, a mixture-of-experts model with 400 billion parameters.
Using a single NVIDIA H100 GPU, researchers demonstrated proof generation at approximately 110 tokens per second, producing proofs around 19 MB in size for a 1,024-token request.
However, generating those proofs required approximately 70 to 95 times the GPU time of ordinary inference using vLLM.
That makes widespread deployment challenging, even though verifying completed proofs can take seconds without specialized hardware.
Interestingly, Claude helped researchers improve the system's performance, while the team used Lean, a formal verification language, to express and check its mathematical security guarantees.
Anthropic has also released executable code, protocol specifications, and verification tools under the Apache 2.0 open-source license.
Topics For More Insights
Anthropic's AI Verification System Faces Limitations Before Real-World Deployment
Despite its promising results, FlashProve isn't ready for widespread commercial use.
The current prototype supports contexts of up to 1,024 tokens and relies on an integer-based model representation that may differ from conventional floating-point inference.
It also protects model weights without concealing architectural details, while its post-hoc verification approach can detect tampering only after the relevant computation.
As reported by TechCrunch, the research remains ongoing, with no announced plans for deployment at Anthropic or another AI laboratory.
Nevertheless, FlashProve demonstrates how cryptographic verification could help customers establish whether the AI model they requested is actually the one serving them.
For now, verifying AI identity remains considerably more expensive than simply trusting the provider.
