{"slug": "google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to", "title": "Google DeepMind’s AlphaProof Code Verifies Signal’s Encryption All the Way Down to Arm Machine Code", "summary": "Google DeepMind announced on November 10 that its AlphaProof Code multi-agent system machine-checked the core of Signal's encryption — PQXDH and the Triple Ratchet — against a compiled 64-bit Arm binary built from libsignal 0.105.4, rather than against source code. A team of five produced roughly 6.5 million lines of Lean in four weeks, establishing forward secrecy, authentication, post-compromise security and handshake secrecy, a task the company compares to efforts like the seL4 microkernel's roughly 20 person-years. DeepMind research scientist Swarat Chaudhuri described it as \"a system that is fundamentally built for human AI collaboration,\" and Signal CTO Ehren Kret said the work expands the scope of earlier verification efforts, with the Lean artifacts published on GitHub.", "body_md": "**Google DeepMind says its AlphaProof Code agents have machine-checked the core of Signal’s encryption, PQXDH and the Triple Ratchet, against the compiled 64-bit Arm binary rather than the source. A team of five did it in four weeks, a job that classically eats years.**\n\nFormal verification has always had a reputation problem. Everyone agrees that a mathematical proof beats a test suite, and almost nobody does it, because the proofs are brutally expensive to write. The seL4 microkernel, the field’s poster child, took around 20 person-years. The CompCert verified C compiler took six years. Google DeepMind now claims it can cut that kind of effort down to weeks, and it picked a high-profile target to make the point: Signal.\n\nOn November 10 the company [introduced AlphaProof Code](https://deepmind.google/blog/introducing-alphaproof-code/), a multi-agent system that writes proofs in the Lean theorem prover. Its first big job covered libsignal, the open-source library at the heart of Signal, and specifically two pieces: PQXDH, the post-quantum key agreement used when two people first start talking, and the Triple Ratchet, the newer scheme that keeps rotating keys as a conversation goes on. According to [The Register](https://www.theregister.com/ai-and-ml/2026/11/10/google-deepmind-melds-humans-and-ai-agents-to-vet-signals-secure-messaging-techniques/5308625), the proofs establish forward secrecy, authentication, post-compromise security and handshake secrecy for that code.\n\n## Why the machine code matters\n\nThe headline detail is what was actually checked. As [CyberInsider](https://cyberinsider.com/google-deepmind-signal-encryption-machine-code-verification/) reports, the target was not Rust source but a 64-bit Arm executable compiled from libsignal 0.105.4. Lean was used to model how those Arm instructions behave, and the security properties were proven against that model. That closes a gap that source-level proofs leave open: a compiler that optimises something in an unexpected way, or a toolchain bug, can quietly break a guarantee you thought you had proven. Here, the thing that was verified is much closer to the thing that actually runs on a phone.\n\nThe scale is the other eye-opener. The agents, which run on what DeepMind calls Gemini Carbon, produced roughly 6.5 million lines of Lean. No human reads that, and nobody has to: Lean’s small trusted kernel checks every step. People still wrote and reviewed the specifications, which is where most of the judgement lives, and an auditing pass flagged parts of the work that needed a human look. DeepMind research scientist Swarat Chaudhuri described it to The Register as “a system that is fundamentally built for human AI collaboration.”\n\nTo show the proofs have teeth, the team also put two previously fixed libsignal bugs back into isolated copies of the code. In both cases the framework produced a machine-checked proof that the modified build violated a security property. That is the right test. A verifier that happily “proves” broken code is worse than no verifier at all.\n\n## What it does not cover\n\nThe usual caveats apply, and they are not small. The proofs assume the underlying cryptographic primitives are perfect. Side channels such as timing and power analysis are out of scope, and so is everything in the Signal app outside these protocol paths, from the UI to the backup code. No new vulnerabilities were reported either, which is good news for Signal users but also means this is a demonstration, not a bug hunt. Signal CTO Ehren Kret said the work expands the scope of earlier verification efforts, and the Lean artifacts are [published on GitHub](https://github.com/google-deepmind/code-verification) for anyone who wants to check them.\n\nThe bigger story is economics. If proving properties of compiled binaries really drops from years to a month for a five-person team, verification stops being something reserved for kernels and aircraft software. Crypto libraries, firmware and hypervisors are the obvious next candidates. The catch is that the specification still has to be right, and that part remains stubbornly human.", "url": "https://wpnews.pro/news/google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to", "canonical_source": "https://hwbusters.com/news/google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to-arm-machine-code/", "published_at": "2026-11-11 16:10:42+00:00", "updated_at": "2026-11-11 16:23:33.186805+00:00", "lang": "en", "topics": ["ai-agents", "ai-research", "ai-safety", "artificial-intelligence"], "entities": ["Google DeepMind", "AlphaProof Code", "Signal", "libsignal", "Lean", "Swarat Chaudhuri", "Ehren Kret", "Gemini Carbon"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to", "markdown": "https://wpnews.pro/news/google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to.md", "text": "https://wpnews.pro/news/google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to.txt", "jsonld": "https://wpnews.pro/news/google-deepminds-alphaproof-code-verifies-signals-encryption-all-the-way-down-to.jsonld"}}