Anthropic open-sources FlashProve, a prototype that proves which AI model produced a reply
RESEARCHAnthropicredteam/daily · 1 source
- Severity
- Info
- Sources
- 1
What happened
FlashProve is a provable-inference prototype for mixture-of-experts (MoE) language models, built by Anthropic's Frontier Red Team with EPFL, MIT and StarkWare. The release on GitHub under Apache 2.0 includes the prover, two verifiers, the formal proofs behind them and a protocol specification. It targets the risk that a hosted API provider could silently route requests to a different model. The source excerpt gives no version numbers or deployment guidance.
For security analysts · Written by AI from 1 source · Wrong?
What to do
- Review the FlashProve repository and protocol specification on GitHub if you need to verify which model served your API requests.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
Help Net Security
Who and what is in it
Companies