Anthropic unveils FlashProve for provable model inference
According to Anthropic, its new open source FlashProve uses cryptographic proofs to verify if a response came from a specified model and is more efficient than prior open source methods, with applications in chat integrity and safety evaluation.
SourceAnalysis
On November 3 2026 Anthropic announced progress on provable inference for large language models with the introduction of FlashProve.
FlashProve builds on established cryptographic techniques to let developers verify whether a given response came from a given AI model and is substantially more efficient than prior open-source methods according to the announcement.
The proofs support applications such as making it harder for attackers to hijack a chat and enabling safety evaluators to confirm that the model under test matches the one served to customers.
Anthropic is open-sourcing FlashProve code on GitHub and has published a full write-up on its research site see Anthropic research page.
Anthropic
@AnthropicAIWe're an AI safety and research company that builds reliable, interpretable, and steerable AI systems.