跳到正文
原文
The Decoder:AI News· Maximilian Schreiner·· 12 天前AI 评分46

Anthropic 发布 FlashProve 研究原型,用密码学核验 AI 回复出自哪个模型且不暴露权重

Anthropic's new prototype proves which model wrote an AI response without exposing its weights

文章要点
  • 1Anthropic 发布 FlashProve 开源原型,用密码学验证 AI 回复确实出自所声称的模型且不暴露权重
  • 2官方称单张 H100 为 400B 参数的 Llama 4 Maverick 生成证明约 110 tokens/秒,单份约 19 MB
  • 3开销约为正常推理的 70 到 95 倍;Anthropic 称此前可比开源系统处理 1000 tokens 要超过 14 小时
  • 4原型仅支持 1024 tokens 上下文,只在模型选单一最可能 token 时有效,暂不适配日常随机输出场景
决策要点

该原型把输出是否真的出自所声称模型的问题,从信任承诺转向密码学核验,直指服务商偷偷换模型和篡改输出这类供应方风险,但离日常可用尚远。

  • 影响:企业目前难以核验服务商是否暗中换用更便宜的模型,只能靠合同约束;对受监管行业,可核验的模型来源会改变审计与合规方式;但 70 到 95 倍开销和 1024 tokens 上限使其短期无法直接采用。
  • 参考:可考虑在模型采购合同中明确版本承诺与替换通知条款;受监管场景值得把可验证推理列入技术选型评估项;技术团队可先在 GitHub 审阅代码,估算未来接入核验的改造成本。
  • 依据:性能与限制数据来自 Anthropic 技术报告,官方口径暂无独立验证;此前可比系统的对比数据也由 Anthropic 提供。
正文

Skip to content

Maximilian Schreiner

Nov 4, 2026

Image description

Anthropic

Anthropic has released FlashProve, a research prototype that uses cryptography to verify whether an AI response actually came from the claimed model. The system still has significant limitations.

Anyone sending a query to an AI model today has no way to verify that the chosen model actually generated the response. A routing service could quietly swap in a cheaper model to cut costs. An attacker could tamper with the data stream. Anthropic says its new open-source research prototype FlashProve is designed to tackle this problem.

FlashProve uses a cryptographic technique called zkSNARK. In simple terms, the model provider first creates a digital fingerprint of the model's weights. With every response, the provider then delivers a compact proof showing that this specific model generated the output. The proof reveals nothing about the secret model weights themselves.

Proof generation is fast enough to be practical

According to the technical report, FlashProve generates proofs at roughly 110 tokens per second on a single H100 GPU for Meta's Llama 4 Maverick, a 400-billion-parameter model. Each proof is about 19 megabytes and can be verified in 17 seconds on a standard CPU. That adds overhead of roughly 70x to 95x compared to normal inference.

That sounds like a lot, but it is a significant jump from earlier systems. The only comparable open-source system previously available took over 14 hours to process 1,000 tokens, Anthropic says. It produced a 93.6-gigabyte proof that took nearly a full day to verify.

Regulators could verify which models providers actually run

Anthropic sees several use cases. Regulators could check whether a provider is actually running only approved models, companies could prove their code was reviewed by a specific safety model and providers could detect tampering on their servers, like an attacker or a more advanced model slipping in unnoticed.

In its Frontier Safety Roadmap, Anthropic had set a goal of building an inference verification prototype by September 2026. The company says it already hit that target through a separate project that signs inference traffic and traces the computation on a trusted server. FlashProve pursues the same goal but takes a purely cryptographic approach and doesn't need trusted hardware.

Context window and sampling limits hold it back

Anthropic stresses that FlashProve is a research prototype. The system currently only proves contexts of up to 1,024 tokens, far below the context windows of current models. It hides the model's weights but not its underlying architecture. And the proof only works when the model picks the single most likely next token, not with the randomized outputs that chatbots typically produce in everyday use.

The full code, Lean proofs, and documentation are available on GitHub under the Apache 2.0 license.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now

  • Full access to every article on THE DECODER
  • No ads
  • Join the comments and community discussions
  • A weekly AI news recap via mail
  • 6x/year: "AI Radar" — deep dives on the AI topics that matter most
  • Daily AI news, always up to date
  • Our full ten-year archive
  • Covered by a team with 10+ years in AI

来源:The Decoder:AI News · the-decoder.com