Anthropic تنشر FlashProve مفتوح المصدر: نموذج أولي يثبت بالتشفير إن الرد طلع فعلًا من النموذج المعلن، مو من نموذج بديلAnthropic releases open-source FlashProve, a prototype that cryptographically proves a reply really came from the stated model, not a substitute

Anthropic has published research on provable inference, along with an open-source system called FlashProve. The problem it tackles: when you talk to an AI model, you have no way to check that the reply came from the model the company says it did, and not from another model a routing service swapped in to save money, or from an attacker who tampered with the response.

The idea is that the provider commits to a specific model, then attaches a cryptographic proof to each reply showing it came from that model, without revealing the model's weights. Anthropic says Claude produced the improvements that make the system much faster than earlier open-source systems, and because Claude's decisions can be hard to evaluate, the team added formal statements in Lean that anyone can audit.

The team tested it on the open Llama 4 Maverick model (400 billion parameters): a single H100 GPU generates proofs at around 110 tokens per second, each proof is about 19 MB per 1,024-token request, and checking it takes about 17 seconds on a CPU with no GPU. For comparison, an earlier open system, VerInf, proved the same model at about 0.02 tokens per second.

But Anthropic calls this intermediate progress and a research prototype, not a finished product: proving takes roughly 70 to 95 times the GPU time of running the model itself, and it currently works only on short contexts of up to 1,024 tokens. The tests were on Llama, not Claude, and the code is released under the Apache 2.0 licence.