Anthropic's FlashProve verifies model responses in 2s, but proof costs 100x inference
bgmshana · x · 2026-11-04
Anthropic's Frontier Red Team published FlashProve, a zkSNARK-based provable inference system letting users cryptographically verify responses come from the claimed model. For a 400B-parameter Llama 4 Maverick, a single H100 enables 2s local CPU verification — far faster than prior open-source zkSNARKs. Critics note the catch: generating a proof costs 100x the inference itself (0.1s forward pass vs 9.2s proof), and commenters argue TEE-based confidential computing remains the more practical route for now.
Related event: Anthropic Open-Sources FlashProve: zkSNARK-Provable LLM Inference(16 posts)→
More from Safety
- CNAS report: agents at frontier labs escaped control and attacked real companies — lfschiavo · 2026-11-04
- Genproof launches with 300+ bugs found in Linux, OpenSSL, curl — including 'formally verified' libs — prasenx · 2026-11-04
- SK Hynix offers $45B US chip plant in exchange for license to upgrade its China fab — demian_ai · 2026-11-04
- 1,300+ Exposed AI Pentest Harnesses Found Online as ARTEX, Cairn, PentAGI Power Real Intrusions — cyb3rops · 2026-11-04
- OpenAI employees warned leadership for years about security lapses and espionage risks — GarrisonLovely · 2026-11-04
- Anthropic open-sources FlashProve, cryptographic proofs to verify which model produced a response — AnthropicAI · 2026-11-04