AlphaProof Code verifies Signal's libsignal in Lean
DeepMind's AlphaProof Code shows one compiled release of Signal's libsignal meets key security properties in one-to-one chats, assuming secure crypto.
What it means for founders
- Specifications are now the expensive part. A team of five finished in four weeks without writing proof code themselves. The paper argues that once proofs are cheap, the hard part becomes choosing what to prove, which makes specification writing and review the opening for developer and security tooling.
- Ask what "verified" covers. When a vendor cites a result like this, ask which properties, which code version and which assumptions it rests on.
- Proof complements bug hunting. Automated, often invalid reports have flooded bug bounties, as Enki covered when Google paused its open source bug bounty, and attackers can move before a fix ships, as with the AI-found PHP flaws. A machine-checked proof rules out whole classes of attack within its scope.
- Watch the open artifacts. DeepMind published the Lean specifications and proofs for outside cryptographers to audit.
What AlphaProof Code proved about Signal
The DeepMind announcement and its technical paper describe proofs about libsignal v0.105.4, Signal's reference implementation, compiled with the standard Rust compiler to AArch64 machine code. The proofs cover the PQXDH handshake and Triple Ratchet in one-to-one sessions. They establish authentication, forward secrecy, handshake secrecy and post-compromise security, with handshake secrecy holding even against a future quantum attacker in the symbolic model.
The agents produced roughly 6.5 million lines of proof in Lean. A reader has to trust about 26,000 lines: the specifications plus a Lean model of AArch64. Lean checks everything else, though the results still rest on Lean itself, the build tooling and the assumed cryptography. To test the specifications, the team restored two previously fixed libsignal bugs and proved each buggy binary violates the symbolic security theorem.
Signal CTO Ehren Kret, quoted in DeepMind's post, said the work gives Signal "even greater confidence that our code delivers on its privacy guarantees."
What the Signal proofs assume and leave out
The proofs treat the low-level cryptographic routines, including SHA-256, AES and Kyber, as contracts and assume those primitives are secure. Side channels, group messaging, sealed sender, server-side protocols and multi-device state are out of scope. The ratchet theorems stop short of proving that SPQR, Signal's newer quantum-resistant ratchet layer, restores security against quantum attackers after a compromise, and the computational proof covers classical attackers only. The paper argues the results carry over to the Android and iOS apps, which run the same Rust code. Strictly, the theorems cover DeepMind's own test build, not Signal's shipped apps.
How AlphaProof Code splits work between people and agents
Humans approved every specification before proving began. According to the paper, agents built on Claude Opus 5.5 and Gemini 4 Carbon drafted and reviewed specifications, and Gemini 4 Carbon agents wrote all of the proofs. A gatekeeper agent rejects any proof unless Lean's kernel verifies it and a comparator tool confirms it matches the approved specification exactly.
What we don't know yet
DeepMind has not said whether AlphaProof Code will be offered outside Google. The authors say they want to verify new libsignal releases before they ship; Signal has not said whether it will. DeepMind has not disclosed costs. The paper says each proof campaign used 40 to 80 agents running 24 to 72 hours on one machine, and a clean proof build could take several hours.
The story
On Tuesday Google DeepMind unveiled AlphaProof Code, a multi-agent framework for formal verification, and used it to prove core security guarantees of compiled Signal protocol code. The result covers specific properties of one libsignal release under stated assumptions, a narrower claim than "Signal is secure."
Sources
Primary sources
Reporting
More in Research
- Podqi finds chatbots sending shoppers to risky sitesA Podqi study found Perplexity, ChatGPT and Gemini showed risky or fake retail sites in 4% to 15% of shopping answers.The Verge · 5d ago

- DeepMind launches AlphaGenome 2 DNA model via API
For founders: A startup cannot use the AlphaGenome Services for company work. The AlphaGenome terms open them only to individuals, non-commercial organizations and…
Google DeepMind · 7d ago - OpenAI claims AI proofs on Kakeya, ETH and QCD

For founders: Watch the ETH review closely. If it holds, problems long assumed hard under ETH may admit faster exact algorithms, though OpenAI itself says practical gains…
OpenAI · 23d ago - GPT-6.1 Sol and Grok 4.7 pick rival parties in study

For founders: API access may not carry the app's guardrails. If your product calls GPT-6.1 Sol or Grok 4.7 directly, a user may get a political recommendation that ChatGPT…
TechCrunch · 25d ago - Microsoft says its AI beat doctors on hard diagnoses

For founders: Orchestration alone is a thin moat. If your health product wraps frontier models in an agent workflow, note how unevenly the orchestrator helped: about 15…
The Verge · 26d ago