Google DeepMind says AI agents formally verified Signal protocol code in four weeks

Google DeepMind says AI agents proved key security properties of the real code behind the Signal protocol. The company announced the work on November 10, 2026, in its own blog post. The framework is called AlphaProof Code.

What happened

According to Google DeepMind's post, researchers guided AI agents to prove security properties of the code that underlies Signal's cryptographic protocol. The protocol is also used in Google Messages.

DeepMind calls AlphaProof Code a multi-LLM agentic framework. The company says this is the first time AI has formally verified a widely deployed cryptographic protocol at this depth. It also says it is the first time such a protocol's implementation has been formally verified by any method.

How the workflow runs

DeepMind describes a split between people and agents.

1. Researchers and a mix of models turn Signal's security goals into formal specifications. Human experts review every one.
2. Agents powered by Gemini Carbon write the proofs in Lean.
3. The agents audit the proofs and flag the parts that need a human look.
4. Lean checks every proof, which verifies the agents' work.

Key details

• The agents generated over six million lines of Lean, according to DeepMind.
• Only one percent of the codebase was flagged for a second look by a human.
• The proofs cover the compiled code, not only the protocol design. One example property is that messages stay protected before and after a device is temporarily compromised.
• DeepMind says verification took four weeks. It says this kind of task historically takes expert teams years.
• As a test, researchers reintroduced two real, previously fixed bugs into an isolated copy of Signal's code. In both cases the system produced a machine-checked proof that the buggy version was insecure.
• DeepMind is open-sourcing its formal verification artifacts and has published a technical paper.

Why it matters

Most protocol verification stops at the design level. DeepMind notes that real-world vulnerabilities often hide in the gap between design and implementation. Proofs that reach the code running on devices target that gap.

Signal's CTO Ehren Kret is quoted in the post saying the approach enables a significant expansion over prior verification work.

These are DeepMind's own claims. DeepMind says scaling the techniques needs more research. It points to pairing them with its CodeMender defense agent.

FAQ

What is AlphaProof Code? It is a Google DeepMind agentic framework for formal verification. AI agents help write a specification, then prove that the implementation meets it, and Lean checks each proof.

Did DeepMind find bugs in Signal? The post does not describe any new bugs. It says the proofs show Signal's core cryptographic code delivers the protections the experts specified. The bug test used two previously fixed bugs that were reintroduced.

Can I look at the verification work? DeepMind says it is open-sourcing its formal verification artifacts and links a technical paper in the post.