Anthropic Releases Open-Source FlashProve for Verifying AI Model Outputs
November 4: The Decoder adds proof generation rate of 110 tokens/sec on H100 GPU, 19 MB proofs verifiable in 17 seconds on CPU, and no trusted hardware needed.
Anthropic open-sources FlashProve, a zkSNARK system proving AI replies derive from committed model weights without exposing them.
Anthropic released FlashProve, a research prototype using zkSNARKs to generate cryptographic proofs confirming that specific replies originated from claimed model weights. The open-source system, developed with EPFL, MIT, and StarkWare, supports local CPU verification and resists impersonation or secret swaps. It was tested on the Llama 4 Maverick 400B parameter Mixture-of-Experts model and enables third-party claims verification such as vulnerability scans. Research is ongoing with no deployment plans.
- Prover, C++ verifier, and Lean verifier released under Apache 2.0
- Tested on Llama 4 Maverick 400B parameter MoE model
- Proof generation: 110 tokens per second on H100 GPU
- Proof size: 19 MB; verification: 17 seconds on 24 CPU threads
- Offline commitment: 39 minutes per model
- 72x overhead compared to vLLM inference
Timeline
dated events in the reporting · oldest first- Nov 3, 2026Anthropic announces FlashProve using zkSNARKs for AI output verification · MarkTechPost
- Nov 3, 2026TechCrunch reports open-source system for provable AI model identity · TechCrunch
- Nov 4, 2026Anthropic open-sources FlashProve prototype tested on Llama 4 Maverick MoE · Help Net Security
- Nov 4, 2026The Decoder reports FlashProve prototype details and limitations · The Decoder
What each outlet adds
beyond the shared facts- TechCrunchOpen-source system resists model impersonation and supports third-party claims verification
- Help Net SecurityTechnical specs including 18.78 MB proof size, 17-second CPU verification, and 72x overhead
- The DecoderPerformance metrics, 1024-token context limit, deterministic outputs, and no trusted hardware requirement
- Vendors
- Anthropic
- Products
- FlashProve
- Organizations
- EPFLMITStarkWareBroad Institute
- AI models
- AnthropicLlama 4 MaverickFlashProve
Coverage
oldest first · one row per article · UTCAnthropic Releases FlashProve for Verifiable AI Replies
Anthropic's FlashProve tool generates zkSNARK proofs to confirm AI responses derive from committed model weights.
Anthropic develops FlashProve for provable AI model identity
An open-source system offering cryptographic proofs that outputs come from claimed models without revealing weights; lightweight and local verification possible.
Anthropic open-sources FlashProve provable inference prototype
FlashProve cryptographic proofs confirm specific models generated specific replies, tested on Llama 4 Maverick MoE model.
Anthropic releases FlashProve for verifying AI model outputs
FlashProve uses zkSNARKs to prove model authorship without exposing weights, limited to 1024 token contexts.