AI agents formally verify Signal protocol code in four weeks

Summary

• Google DeepMind says AI agents mathematically proved key security properties of Signal protocol's actual code.
• Formal verification, the highest security assurance standard, historically takes expert teams years of manual work.
• It is the first widely deployed cryptographic protocol implementation verified this deeply, by any method.

The latest

Key security guarantees of the code behind the Signal protocol have been mathematically proven by AI agents, Google DeepMind said, in what it describes as the first formal verification of a widely deployed cryptographic protocol implementation at this depth. The protocol secures conversations in apps including Signal and Google Messages. The process took four weeks.

Details

• The system: AlphaProof Code is described by Google DeepMind as a multi-LLM agentic framework built for formal verification, pairing human expertise with AI agents. Researchers worked with a mix of models to turn Signal's stated security goals into formal mathematical specifications, with human experts reviewing each specification before proofs began.
• The scale: Agents powered by Gemini Carbon generated more than 6 million lines of Lean, the language used to mathematically prove code correctness, according to the company. The agents then audited their own proof and flagged sections needing a second human look — about 1% of the codebase.
• What was proven: Google DeepMind said the agents proved the compiled code satisfies key security properties down to the exact code running on devices, including keeping messages protected before and after a device is temporarily compromised. The company did not detail the full list of properties covered.
• The timeline: Verification at this complexity took four weeks, according to the company, against a task it says historically takes expert teams years. Formal verification — proving mathematically that code runs without flaws — is presented as the highest standard of security assurance, but remains difficult and labor-intensive.
• The stress test: To check the system could separate sound code from flawed code, researchers reinserted two real, previously fixed bugs into an isolated version of Signal's codebase. In both cases Google DeepMind said AlphaProof Code produced a machine-checked proof that the modified version was insecure.
• Signal's position: Ehren Kret, Chief Technology Officer of Signal, said translating live code into valid Lean theorems using AI agents is a major breakthrough, and that rigorous verification of the implementation strengthens confidence in the privacy guarantees of Signal's code as AI-driven security threats grow.
• The release: Google DeepMind said it is open-sourcing the formal verification artifacts for community use. The announcement did not name a distribution date, licence terms, or which other protocols or systems the framework will be pointed at next.
• The authors: The announcement was published on the Google DeepMind blog on November 10, 2026, under the names of Pushmeet Kohli and Swarat Chaudhuri, under the title introducing AlphaProof Code as a breakthrough in verifying complex software at depth.

Background

The Signal protocol is the encryption layer used by Signal and Google Messages and underpins private messaging for billions of users. Verifying a protocol's design is common; verifying the implementation — the code actually compiled and shipped to devices — has been the harder, rarely attempted step.

Between the lines

The claimed advance is less the proof itself than the cost curve: four weeks against years, with humans reviewing roughly 1% of the codebase, is what would make formal verification practical beyond research settings. The reinserted-bug test is the sharpest evidence offered, since a system that proves everything secure proves nothing. Independent scrutiny becomes possible once the verification artifacts are released openly.

What's next

Watch for the open-sourced verification artifacts and independent review of the Lean proofs, any disclosure of licence terms or release timing, and whether AlphaProof Code is applied to other deployed cryptographic systems.

Source: Google DeepMind