Thursday 12 November 2026 Archive edition · go to today's edition NL EN FR
← Tech · AI & LLMs
Tech · AI & LLMsFormal verification

Four weeks and millions of lines of maths: Google DeepMind's AI proves Signal's security right down to machine code

AlphaProof Code showed that the core of the messaging app does what it promises, although the app itself and timing attacks remain out of scope.

· latest source update di · 3 sources · Hardware Busters, AlphaSignal, Qodo
Illustration for this article

Google DeepMind on Tuesday presented AlphaProof Code, a system of cooperating AI agents that writes mathematical proofs in the Lean proof assistant. Its first major assignment was libsignal, the open source library that handles Signal's encryption. According to Hardware Busters, a team of five people worked on it for four weeks, with the agents producing about 6.5 million lines of Lean. AlphaSignal and Qodo put it at more than six million lines.

Notably, it was not the Rust source code that was checked but the compiled program for 64-bit Arm processors, in other words the code that actually runs on a phone. Properties proven include forward secrecy, authentication and recovery after a compromised key, for both the PQXDH key exchange and the Triple Ratchet that keeps generating new keys during a conversation. Human experts wrote and approved the specifications, and according to Qodo the agents flagged about 1 percent of the work for an extra human check. The rest was checked mechanically by Lean.

To test the system, the team reinserted two previously fixed bugs into an isolated copy of libsignal. In both cases, AlphaProof Code produced a proof that this version violated a security property. No new flaws were found. The proofs assume that the underlying cryptography is sound and say nothing about the rest of the app or about attacks based on timing or power measurements. The Lean files are public on GitHub, but according to AlphaSignal the tool itself is not freely available. Google declined to say what the analysis cost.

Spotted a mistake in this article? Let us know.