Overview
- Anthropic’s Frontier Red Team published FlashProve as open‑source research software that uses zkSNARK cryptography to prove a given output was produced by a committed model without revealing model weights, a move announced on Tuesday to address gaps in API trust.
- Benchmarks show FlashProve can generate roughly 110 prompt tokens per second on a single NVIDIA H100, produce a proof of about 18.78–19 MB, and verify that proof in roughly 17–21 seconds on CPU threads.
- Generating proofs is far more expensive than standard inference, costing about 70–95 times more GPU time than a plain vLLM run, and proofs are made in post‑hoc batches rather than inline during streaming responses.
- The prototype has strict technical limits: it proves only 1,024‑token windows, uses integer inference and greedy decoding instead of floating‑point sampling, is not integrated into the Claude API, and the GitHub release is labeled unmaintained.
- Anthropic provides a Lean formal specification that proves key security properties, while faster C++ and GPU implementations exist but are not themselves formally verified, so the tool is currently suited to audits and research rather than routine real‑time API use.