DeepMind agents formally verify Signal's encryption code with 3M+ lines of proofs

Google DeepMind announced that its AI agents have formally verified the code behind Signal's encryption. Using agents powered by Gemini 4 Carbon, the team proved key security properties of the machine code, producing more than 3 million lines of proofs. DeepMind's Pushmeet Kohli described it as a major milestone for his team, and the announcement spread widely on X.

Formal verification is different from testing. Testing checks specific inputs; formal verification proves mathematically that code behaves correctly for every possible input, within a defined model. It is the strongest assurance software can get, but it has historically been extremely expensive. Verifying even small systems has taken expert teams years. Working at the machine-code level adds rigor because it rules out bugs introduced by compilers, though it makes the proofs much larger. AI agents that can write and check proofs could change those economics.

This is one of the clearest examples yet of AI improving software security rather than threatening it. Much of this year's security conversation has focused on models finding vulnerabilities: VulnCheck's tracking of Anthropic-discovered vulnerabilities was a top HN discussion this week, and labs are marketing models for cyber capabilities. Proving the absence of whole classes of bugs is a defensive counterweight. It also gives Google a strong real-world showcase for Gemini 4 Carbon's long-horizon reasoning, at a time when testers are criticizing the model's latency.

There are limits. Researchers noted the sheer scale required: 3 million lines of proofs is not something most projects can produce or review. A proof is only as strong as its specification; if the properties being proven miss a real-world attack, the code can still be vulnerable. Verification also covers specific code at a point in time and must be redone as the code changes. Watch whether DeepMind releases the tooling, whether other critical open-source libraries get similar treatment, and how Signal integrates verification into its development process.