Story
Nov 11, 2026
DeepMind’s Signal proof promises safer code—but not yet unhackable software
Google DeepMind says its AI agents formally verified key protections in Signal’s production cryptographic code in four weeks. The result is a major confidence boost, though researchers concede that scaling such proofs across all software remains a formidable challenge.
Google DeepMind announced a formal-verification result for Signal’s production cryptographic implementation, arguing that AI can now close a perilous gap between a protocol’s design and the machine code that actually runs on devices. The company says AlphaProof Code, using Gemini Carbon-powered agents under human review, proved selected security properties of the core code in four weeks—a process it says would historically have taken expert teams years.1
The timing matters because cryptographic protocols safeguard everything from private messages to banking transactions, and weaknesses often emerge during implementation rather than in the abstract design. In its announcement, DeepMind said agents generated more than six million lines of Lean proof code, while human experts set the goals and reviewed the results. It also reintroduced two old, fixed Signal bugs; the system correctly showed both altered versions were insecure.1
Signal welcomed the result as more than a laboratory exercise. Chief technology officer Ehren Kret said the approach enables “significant expansion” beyond earlier, narrower verification efforts and gives the organization “even greater confidence” in Signal’s privacy guarantees.2 DeepMind chief executive Demis Hassabis amplified the milestone on X, describing it as proof of key security properties in the machine code behind Signal’s messaging protocol.
3
Google CEO Sundar Pichai stressed that the work verified the protocol down to compiled device code—where theoretical assurances can otherwise fail to survive contact with real software.
4
But the breakthrough is not a declaration that Signal, or software generally, is invulnerable. Reporting on the project notes that DeepMind did not formally verify the entire Signal mobile app, and that the central question is whether AI-assisted verification can be extended across vast codebases at a workable cost. Research scientist Swarat Chaudhuri put the limitation plainly: “there is a long distance to go before we are able to scale this to all the software of the world.”2
For now, the proof is a sharp demonstration of what AI-assisted security review can do—not yet the universal proof that every program is safe.