Google DeepMind: AlphaProof Code Verifies Signal Cryptographic Implementation

Google DeepMind researchers have guided AI agents to mathematically prove key security properties of the actual compiled code underlying the Signal protocol, according to a November 10, 2026 research post by Pushmeet Kohli and Swarat Chaudhuri. The work used AlphaProof Code, described as a novel agentic framework for formal verification that combines human expertise with AI agents. The research team says this is the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth, and the first time such an implementation has been formally verified by any method, with or without AI. Proofs were generated in Lean and checked by the Lean proof assistant.

Key Takeaways

• AlphaProof Code is a multi-LLM agentic framework that splits verification work between human experts, who work with AI to decide what to prove, and a group of AI agents that produce the proofs.
• Agents powered by Gemini Carbon generated over six million lines of Lean, and the agents then audited the proof and flagged one percent of the codebase for a second human look.
• The research team says its agents proved that the compiled code fulfills key security properties, such as keeping messages protected before and after a device is temporarily compromised, down to the exact code running on devices.
• The verification process took four weeks, a task the team says historically takes expert teams years.
• To test detection, researchers reintroduced two real, previously fixed bugs into an isolated version of Signal's code; the system produced a machine-checked proof that the buggy version was insecure in both cases.

Context & Analysis

Real-world vulnerabilities often hide in the gap between design and implementation, and bridging that gap makes proofs vastly more complex. Google DeepMind describes AlphaProof Code as combining human expertise with AI agents. Researchers collaborated with a mix of models to translate Signal's security goals into formal mathematical specifications, with human experts reviewing every specification. Agents powered by Gemini Carbon then did the proofs, generating over six million lines of Lean, a computer language that mathematically proves the accuracy of code.

After proof generation, the agents audited the proof and highlighted areas needing a second look by a human, just one percent of the codebase. Because Lean automatically checks the rest, the team says its experts could focus review where human judgment matters most. The distinction between a Lean machine-checked proof and an LLM assertion is central to how the claim should be read: in this framework, AI agents help users write a specification and then autonomously prove the implementation meets it, but every proof is checked by Lean to verify the agents' work. The verification claim therefore rests on the Lean checker, not on model output alone. The scope is Signal's core cryptographic implementation and the properties specified by experts; it is not proof that all Signal app software is bug-free or immune to compromise.

Competitive Landscape

The comparison is limited to documented components and roles within this research programme. The table below compares documented roles and approaches from the source, bounded strictly to what is stated.

What It Means

For Enterprise Buyers

For organisations evaluating security assurance for cryptographic or other safety-critical software, this research illustrates a possible procurement consideration: formal verification can target compiled code rather than only protocol design. Buyers weighing such work should ask whether proofs are Lean machine-checked, what properties were specified, which code version and build were covered, and how human review was scoped. Because the evidence covers a research project on Signal's core cryptography and states that scaling the techniques requires continued research, buyers should not treat it as evidence of a commercial product, universal security or cost savings.

For Investors

The announcement is a research milestone, not a disclosed commercial offering or revenue event. Investors evaluating AI-assisted security assurance should separately investigate market demand, adoption, pricing and financial outcomes. Investors should treat any commercial implications as conditional on future productisation, independent validation and broader deployment, rather than an assumed result of this research.

Forward Outlook

Google DeepMind says verifying Signal's core cryptography lays the groundwork to secure other vital systems, from medical device firmware to power grid software, and that scaling these techniques requires continued research. The team is open-sourcing its formal verification artifacts for community use. Milestones to watch, per the source, include broader application of the approach and pairing with automated defence agents such as CodeMender to counter modern cyber threats.

Risks and unanswered questions remain bounded by the evidence: the proof covers specified security properties of Signal's core cryptographic implementation, not every aspect of the apps that use the protocol; the two-bug test was conducted on an isolated version of the code; and readers should seek independent validation, costs and timelines before assuming readiness for other systems. Whether the method generalises beyond Signal, and how human review scales as codebases grow, are open questions the source does not resolve.

Direct-source note: This article is based solely on the Google DeepMind research post. See Introducing AlphaProof Code.