Anthropic's FlashProve Wants a Receipt for Every AI Answer — Explained Honestly
When you ask an AI model a question through an API, you are taking something on trust. The provider says the answer came from the model you paid for — the big, capable one, not a cheaper stand-in quietly swapped in behind the scenes. You have no way to check. Anthropic's Frontier Red Team has now published a research prototype that tries to change that: FlashProve, an open-source system for…
Anthropic's FlashProve is a research prototype aimed at providing cryptographic proof of the specific model weights that generated an AI model's response. This proof system, based on zkSNARKs, allows verification of whether a specific set of model weights produced a given set of tokens, without revealing the actual model weights.
The prototype works on Meta's Llama 4 Maverick, a 400-billion-parameter model, using 4-bit quantized weights and greedy decoding. It is currently a research release, not a production product, and only supports integer weights under greedy decoding for contexts up to 1,024 tokens. FlashProve is targeted at catching potential issues such as a provider substituting a cheaper model or tampering with the token stream.
However, it is not a fully deployed system and has significant limitations, including high proving costs and limited support for advanced model architectures.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
