As cited
Copy frozen at (site build).
ai security
Anthropic open-sources FlashProve, its provable inference prototype for MoE models
Anthropic released FlashProve, an open-source prototype that generates cryptographic proofs confirming which language model generated a specific response to a given prompt. Developed with collaborators from EPFL, MIT, and StarkWare, the tool addresses verification challenges when using hosted model APIs by providing formal proofs and a protocol specification under Apache 2.0 licensing.
Why it matters: Security practitioners and application programming interface (API) users need assurance that hosted models are performing as claimed; this prototype enables cryptographic verification of model identity and output authenticity for mixtue-of-experts (MoE) architectures.
- Source published
- First seen by Cybersecurity Tracker