Anthropic open-sources FlashProve for verifiable LLM inference
The cryptographic tool generates proofs to verify whether a specific model produced a response, running more efficiently than prior open-source methods.
Key points
- The code is available on GitHub under anthropics/FlashProve alongside a research write-up.
- The tool is designed to prevent chat hijacking.
- Safety evaluators can use the proofs to confirm that models served to customers match tested versions.
Why it matters
Developers and evaluators can cryptographically confirm that model outputs come from the claimed model rather than a tampered or hijacked system.
Sources
- Anthropic Introduces FlashProve for Verifiable LLM Inference (Anthropic, Nov 3, 2026, 8:15 PM)