Anthropic and peers advance provable inference with zkSNARKs

Researchers from Anthropic, EPFL and others use zkSNARKs to cryptographically prove that responses came from the claimed model, guarding against substituted or tampered outputs. Their FlashProve system generates proofs on Llama 4 Maverick (400B parameters) at about 110 tokens/second on one H100, with proofs of ~19 MB per 1,024 tokens and ~70-95x the GPU time of plain inference.

• Anthropic unveils FlashProve to verify AI model identity2026-11-03
• MIT study finds AI agents play spontaneously on virtual island2026-10-26
• Enterprises need artificial specific intelligence, not AGI2026-10-22
• AI takeover debate resurfaces as Anthropic researcher resigns2026-11-04
• MIT study: GPT-6 Astra can rewrite other models' weights2026-11-04