Google 新 AI 工具证明 Signal 加密确实有效

NurPhoto / Contributor,来自 Getty Images
研究人员正式验证了 Signal 加密代码的一部分确实能正常工作。
长期以来,对代码进行形式化验证既困难又耗时。
AlphaProof Code 是朝着更强安全验证迈出的一步。
Google DeepMind 的研究人员利用 AI 对 Signal 底层代码的部分内容进行了形式化验证——这意味着他们已从数学上证明,这款加密通信应用的代码至少有一部分确实如其所述那样运行、没有错误。鉴于近来 AI 系统在软件中发现了大量漏洞,而又有如此多人的通信由 Signal 的密码学技术保护,这一点令人安心。WhatsApp 和 Google Messages 使用的是同一套加密技术。
DeepMind 如何验证 Signal
近来,OpenAI 等 AI 公司接连取得数学突破,某些情况下用 AI 解决了已近百年历史的数学难题。通过如此大规模的计算,AI 模型可以大幅加速数学家所做的许多人工工作。
(披露:ZDNET 的母公司 Ziff Davis 于 2025 年 4 月对 OpenAI 提起诉讼,指控其在训练和运行 AI 系统时侵犯了 Ziff Davis 的版权。)
借助 AlphaProof Code,并在 Signal 基金会此前验证工作的基础上,DeepMind 的研究人员把验证任务分成两部分:人类与智能体协作,确定需要证明什么;随后由多个运行该公司最新模型 Gemini Carbon 的智能体来完成证明。研究人员写道,这些智能体还审计了证明,并“标出了需要人类再检查一遍的部分——仅占代码库的 1%,极大节省了时间”。
为进一步检验,研究人员还测试了 AlphaProof Code 区分正确代码与有缺陷代码的能力:他们向它展示了两个此前已修复的 Signal 漏洞,它都给出了正确的证明。
研究人员写道:“我们的工作标志着,AI 首次被用于如此深入地形式化验证一个广泛部署的加密协议。”“这也是此类广泛部署协议的实现首次通过任何方法被形式化验证——无论是否借助 AI。”
通常,研究人员需要手动构建证明来验证软件,耗时要长得多。这一劳动密集型的门槛,让形式化验证长期像是一片遥不可及的应许之地,直到现在。DeepMind 团队的验证工作用一个月完成的事,通常可能需要专家花费数年。
DeepMind 并未验证 Signal 的全部代码,但指出,在 AI 正在颠覆整个网络安全领域之际,这次测试是迈向更完善安全基础设施的一步。
研究人员写道:“借助自动定理证明对 Signal 核心协议进行严格验证,让我们更有信心:我们的代码能够兑现其隐私承诺。”
这一突破再次说明,AI 既是对网络安全的巨大风险,也是同样强大的解决方案。2026 年尤其以 AI 相关的安全事件为标志,其中许多事件都涉及智能体。与此同时,Anthropic、Google 和 OpenAI 等公司正在开发专门面向网络安全的模型,包括 Anthropic 的 Claude Mythos。
DeepMind 并未透露将其研究成果作为消费级产品发布的任何计划,不过研究人员正在开源其代码,以帮助他人开展形式化验证工作。DeepMind 补充道:“验证 Signal 的核心加密技术,为保护从医疗设备固件到电网软件等其他关键系统奠定了基础。”
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Google’s new AI tool just proved Signal encryption really works – here’s how)
NurPhoto / Contributor vis Getty Images
Researchers formally verified that a portion of Signal’s encryption code works.
Formally verifying code has historically been hard and time-consuming
AlphaProof Code is a step toward stronger security verification.
Researchers at Google DeepMind used AI to formally verify portions of Signal‘s underlying code — meaning they’ve mathematically proven that at least parts of the encrypted messaging app’s code do what they say, without errors. That’s reassuring given how many vulnerabilities AI systems have discovered in software lately, and how many people’s communications are protected by Signal’s cryptography. WhatsApp and Google Messages use the same encryption.
How DeepMind verified Signal
AI companies like OpenAI have been on a math-breakthrough tear as of late, in some cases solving nearly century-old math problems with AI. By computing at such vast scale, AI models can rapidly accelerate much of the manual work mathematicians do.
(Disclosure: Ziff Davis, ZDNET’s parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
Using AlphaProof Code and building on earlier verification work by the Signal Foundation, DeepMind researchers divided the verification task into two camps. Humans worked with agents to identify what they needed to prove, and multiple agents running Gemini Carbon, the company’s latest model, then executed it. The agents also audited the proof and “highlighted areas that needed a second look by a human — just 1% of the codebase, a massive time savings,” the researchers wrote.
For good measure, researchers tested AlphaProof Code’s ability to distinguish correct from flawed code by showing it two previously fixed Signal bugs, for which it produced correct proofs.
“Our work marks the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth,” the researchers wrote. “It’s also the first time the implementation of a widely deployed protocol like this has been formally verified by any method, with or without AI.”
Normally, researchers need to manually develop a proof to verify software, which takes much longer. That labor-intensive barrier has made formal verification a distant promised land, until now. The DeepMind team’s verification took a month to do what might normally take experts years.
DeepMind did not verify all of Signal’s code, but noted that this test is a step toward improved security infrastructure at a time when AI is disrupting the entire cybersecurity space.
“Having the core Signal protocol rigorously verified with automated theorem proving gives us even greater confidence that our code delivers on its privacy guarantees,” the researchers wrote.
The breakthrough reiterates how deeply AI is both a massive risk for cybersecurity and an equally powerful fix. The year 2026 has been especially marked by AI-related security incidents, many of which involve agents. At the same time, companies like Anthropic, Google, and OpenAI are developing models geared specifically toward cybersecurity, including Anthropic’s Claude Mythos.
DeepMind did not share any plans to release its findings as a consumer product, though the researchers are open-sourcing their code to help others with formal verification efforts. “Verifying Signal’s core cryptography lays the groundwork to secure other vital systems, from medical device firmware to power grid software,” DeepMind added
Please accept cookies to view the comments powered by Disqus.
Radhika Rajkumar is a senior editor at ZDNET based in New York City. She covers AI, specializing in safety, privacy and security, policy, education, and synthetic media. She also leads ZDNET's newsletter strategy. Radhika holds a Masters in Creative Publishing and Critical Journalism from The New School. See full bio
原文为英文,正文由自动抓取并保留原文。
原文发布于 ZDNet(Google’s new AI tool just proved Signal encryption really works – here’s how),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。