[{"data":1,"prerenderedAt":14},["ShallowReactive",2],{"artikel-en-berita-deepmind-alphaproof-code-signal-11-november-2026":3},{"slug":4,"lang":5,"title":6,"description":7,"date":8,"updated":9,"topic":10,"product":9,"cover":11,"readMin":12,"html":13},"berita-deepmind-alphaproof-code-signal-11-november-2026","en","DeepMind proves the Signal protocol's code is secure with AI agents","Google DeepMind announced AlphaProof Code, an AI framework that mathematically proves the Signal protocol's real code meets its security guarantees.","2026-11-12",null,"berita","\u002Fimages\u002Fartikel\u002Fberita-deepmind-alphaproof-code-signal-11-november-2026.webp",4,"\u003Cp>Google DeepMind announced AlphaProof Code on 10 November 2026, an AI framework that mathematically proved the actual code behind the Signal protocol meets its key security guarantees. The protocol protects billions of private conversations in apps like Signal and Google Messages, and is also used in WhatsApp.\u003C\u002Fp>\n\u003Cp>It is the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth, and the first time any method, with or without AI, has verified the real implementation of a protocol like this.\u003C\u002Fp>\n\u003Ch2>1. What AlphaProof Code is\u003C\u002Fh2>\n\u003Cp>AlphaProof Code is a multi-agent framework that splits formal verification work the way a research team would. Humans work with AI to decide what to prove, and a group of AI agents then does the proving.\u003C\u002Fp>\n\u003Cp>Researchers translated Signal&#39;s security goals into formal mathematical specifications in the Lean language. Every specification was reviewed by a human before it was accepted. According to The Register, the work verified the post-quantum PQXDH handshake and the Triple Ratchet, two core components of the Signal protocol.\u003C\u002Fp>\n\u003Ch2>2. How the agents did the proofs\u003C\u002Fh2>\n\u003Cp>Once the specification was approved, agents powered by Gemini Carbon generated all of the Lean proofs. DeepMind&#39;s blog puts the total at over six million lines, while SecurityWeek and The Register report a figure of around 6.5 million lines. Lean automatically rejects any invalid proof.\u003C\u002Fp>\n\u003Cp>The agents then audited their own proofs and flagged only about one percent for human review. Because Lean checks the rest automatically, the experts could focus their judgment where it matters most.\u003C\u002Fp>\n\u003Cfigure>\n  \u003Cimg src=\"\u002Fimages\u002Fartikel\u002Fberita-deepmind-alphaproof-code-signal-11-november-2026\u002Falphaproof-code-aliran-kerja.webp\" alt=\"Workflow diagram of AlphaProof Code: humans approve a Lean specification, Signal's Rust code is compiled, and a team of prover agents submits proofs to a trusted gatekeeper\" loading=\"lazy\" decoding=\"async\" width=\"950\" height=\"499\">\n  \u003Cfigcaption>The AlphaProof Code workflow: humans approve the specification, agents generate proofs, and a trusted gatekeeper accepts or rejects them. Image: Google DeepMind.\u003C\u002Ffigcaption>\n\u003C\u002Ffigure>\u003Ch2>3. Proofs down to machine code\u003C\u002Fh2>\n\u003Cp>Most protocol verification stops at the design level. DeepMind went further: the proofs show that compiled code satisfies the key security properties, including keeping messages protected before and after a device is temporarily compromised, down to the machine code that runs on real devices.\u003C\u002Fp>\n\u003Cp>According to SecurityWeek, the proven guarantees include forward secrecy, post-compromise security and authentication. Even at this complexity, verification took just four weeks of work by five researchers, a task that historically takes expert teams years.\u003C\u002Fp>\n\u003Ch2>4. Tested against old bugs\u003C\u002Fh2>\n\u003Cp>To confirm the system can tell correct code from buggy code, the team reintroduced two old, already-fixed Signal bugs into an isolated version of the code. In both cases, AlphaProof Code produced a machine-checked proof that the buggy version was insecure.\u003C\u002Fp>\n\u003Cp>Wired reports the analysis found no new vulnerabilities in the portions of the protocol it examined. The researchers have also released their specification and proofs as open-source artifacts for the community.\u003C\u002Fp>\n\u003Ch2>5. The limits\u003C\u002Fh2>\n\u003Cp>The proofs have boundaries. According to SecurityWeek, the results cover Google&#39;s own build of the code, run through a small command-line wrapper, and not Signal&#39;s own applications. Side channels were excluded from parts of the analysis. Wired stresses that the full Signal mobile app was not verified.\u003C\u002Fp>\n\u003Cp>The reaction was positive. Ehren Kret, Signal&#39;s chief technology officer, said using AI agents to translate live code into valid Lean theorems is a major breakthrough, and that it adds confidence in Signal&#39;s privacy guarantees.\u003C\u002Fp>\n\u003Ch2>6. Gemini 4 Carbon opens up\u003C\u002Fh2>\n\u003Cp>The announcement came with more news. Google also expanded access to Gemini 4 Carbon for paying developers and businesses via the Gemini API, at introductory rates of 2 dollars per million input tokens and 10 dollars per million output tokens until 30 November 2026, after which the rates double.\u003C\u002Fp>\n\u003Cp>Google also said Wiz used Carbon to find zero-days at two major SaaS providers within two hours, and its Cloud Vulnerability Research team used the model to confirm that two bugs in a popular open source project were exploitable.\u003C\u002Fp>\n\u003Ch2>What it means for you\u003C\u002Fh2>\n\u003Cp>Formal verification has long been too expensive for most software. If AI can cut the process from years to weeks, the highest standard of security assurance can reach more of the systems you use every day.\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>Your messages stay protected.\u003C\u002Fstrong> The protocol behind WhatsApp and Google Messages conversations is now verified down to machine code, not just at the design level.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>No new vulnerabilities were found.\u003C\u002Fstrong> This analysis is extra independent validation, not a security warning, so there is nothing you need to do.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>AI is becoming a security checking tool.\u003C\u002Fstrong> The same pattern, humans set the goals and agents do the heavy technical work, could spread to banks and critical infrastructure.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>The proofs are public.\u003C\u002Fstrong> The specification and proofs are openly released, so other system designers can adopt the method.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Watch for similar tools.\u003C\u002Fstrong> If you build software, AI-assisted verification frameworks may become a practical option in the coming years.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2>Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fintroducing-alphaproof-code\u002F\" target=\"_blank\" rel=\"noopener\">Introducing AlphaProof Code: a breakthrough in verifying complex software at depth\u003C\u002Fa> (Google DeepMind)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.wired.com\u002Fstory\u002Fa-new-ai-tool-from-google-deepmind-formally-verified-the-signal-protocol\u002F\" target=\"_blank\" rel=\"noopener\">A New AI Tool From Google DeepMind Proved the Signal Protocol Is Secure\u003C\u002Fa> (Wired)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.securityweek.com\u002Fgoogle-widens-gemini-4-carbon-access-after-ai-agents-verify-signal-code\u002F\" target=\"_blank\" rel=\"noopener\">Google Widens Gemini 4 Carbon Access After AI Agents Verify Signal Code\u003C\u002Fa> (SecurityWeek)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.theregister.com\u002Fai-and-ml\u002F2026\u002F11\u002F10\u002Fgoogle-deepmind-melds-humans-and-ai-agents-to-vet-signals-secure-messaging-techniques\u002F5308625\" target=\"_blank\" rel=\"noopener\">Google DeepMind melds humans and AI agents to vet Signal&#39;s secure messaging techniques\u003C\u002Fa> (The Register)\u003C\u002Fli>\n\u003C\u002Ful>\n",1795068892816]