AlphaProof Code

What it is

Formal verification is a method of mathematically proving that code works correctly without defects. It is considered the highest level of security assurance, but it is difficult and labor-intensive. AlphaProof Code is a multi-LLM agent framework that divides this work up the way a research team would. On November 10, 2026, Google DeepMind announced that it had used the framework to prove that the actual code of the Signal Protocol meets key security properties. The Signal Protocol is the encryption technology that protects private conversations in messengers such as Signal and Google Messages.

How it works

First, human experts work with several AI models to translate Signal's security goals into a formal mathematical specification (a precise statement of what is to be proved), and humans review that specification item by item. Next, agents based on Gemini Carbon produced more than 6 million lines of proof in a language called Lean. Lean is a language in which a computer automatically checks whether a proof is correct, so an incorrect proof written by AI will not pass. Finally, agents audited the proofs and flagged the parts humans needed to look at again, which reportedly covered 1% of the codebase.

Why it matters

Until now, verification of cryptographic protocols has mostly stopped at the design stage. Real vulnerabilities often hide in the gap between design and implementation, but this time verification extended to the compiled code that actually runs on devices. The verified properties include one that protects messages before and after a device is temporarily compromised. Google DeepMind said it finished in four weeks work that traditionally took expert teams several years, and that this is the first time a widely used protocol has been formally verified to this depth, with or without AI. In an experiment where two real, already-fixed bugs were reintroduced, it reportedly produced proofs that the code was insecure in both cases.

Things to know

A proof guarantees only against the specification humans have defined. It does not mean risks outside the specification are absent, so the human review of the specification remains essential. The results and the "first" claim rest mostly on Google DeepMind's announcement, so external review is worth watching. The company also said more research is needed to scale this further. The formal verification artifacts have been released as open source and a technical paper is available, so others can check the results.

Example

The story on this site, 'Google DeepMind releases AlphaProof Code: AI agents formally verify Signal's core encryption code down to machine code with Lean proofs', is a representative example. The researchers compiled libsignal, the reference implementation of the Signal Protocol, to machine code and then had agents prove in Lean that the code upholds the specified security properties. Because Lean automatically checks the rest, the experts focused only on the parts the agents flagged. Google DeepMind expects this approach could be used to protect critical systems such as medical device firmware and power grid software.

Where it came up on this site

• briefing Google DeepMind releases 'AlphaProof Code': AI agents formally verify Signal's core encryption code down to machine code with Lean proofs 2026-11-10

Related terms

AI agent Multi-agent Security AI Human-in-the-loop Coding agent LLM

References

• AlphaProof Code - Formal verification of Signal's encryption implementation at the machine-code level (GeekNews)
• Introducing AlphaProof Code: a breakthrough in verifying complex software at depth (Google DeepMind)
• Google DeepMind Used AlphaProof Code to Analyze the Signal Protocol (Mezha.net)
• Google DeepMind Unveils AlphaProof Code, Revolutionizing Software Verification with AI (The Chenab Times)
• DeepMind’s AlphaProof Code Verifies Signal Protocol Encryption Guarantees (JournalArta)
• AlphaProof Code: Signal Verified at the Machine Level (OrcaRouter)