Google DeepMind's research team has completed a formal verification of the security of Signal messages, claiming it is the first time artificial intelligence has assisted in constructing a proof that corresponds the security guarantees at the protocol level with a widely used cryptographic implementation.
This work was accomplished using the AlphaProof Code, a multi-agent framework for software formal specification and verification. The process began with human-guided agents writing Lean specifications and assessing the semantics of machine code, followed by agents generating proofs which were then checked against the specifications by the Lean kernel.
The verified subjects were Signal's post-quantum PQXDH handshake and Triple Ratchet, confirming the design of libsignal in terms of forward secrecy, authentication, recovery from key leakage, and handshake confidentiality. This brings the initiative proposed by the Beneficial AI Foundation in April of this year one step closer: the foundation aims to use Lean formal verification for the Signal protocol and applications.
The specific approach involved collaborating with multiple models to translate Signal's security objectives into mathematically rigorous specifications reviewed by humans, followed by the Gemini Carbon agent generating Lean code based on these specifications to prove the correctness of the libsignal implementation. The agents generated a total of 6.5 million lines of code, after which the team conducted an audit to highlight parts requiring human review.
Google DeepMind Successfully Verifies Signal Message Security
“This is a system fundamentally built for collaboration between humans and artificial intelligence,” said Swarat Chaudhuri, a research scientist at Google DeepMind and a professor at the University of Texas at Austin. He noted that the traditional approach involves mathematicians defining clear problems for AI to prove, whereas this time the focus was first on determining what needed to be proven.

