TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Google DeepMind says its AlphaProof Code framework used AI agents and Lean, a proof-checking language, to formally verify specified security properties of code underlying the Signal protocol. The company reports that the work took four weeks and that it is releasing verification artifacts; independent review and the scope of the proofs remain important questions.
Google DeepMind said on November 10, 2026, that researchers used its AlphaProof Code framework to mathematically prove specified security properties of code underlying the Signal protocol, which protects private conversations in apps including Signal and Google Messages. The company says the work reaches from security goals to compiled code running on devices, addressing a difficult gap between a protocol’s design and its implementation.
DeepMind describes AlphaProof Code as a multi-model, agentic framework that combines human decisions about what to prove with AI agents that develop formal proofs. Researchers worked with several models to translate Signal’s security goals into mathematical specifications, which human experts reviewed. Agents powered by Gemini Carbon then generated more than six million lines of Lean, a language whose proof checker validates mathematical proofs.
According to DeepMind, the framework flagged about 1% of the codebase for additional human review, while Lean checked the rest of the proof material. The company says the verification took four weeks. That timing is DeepMind’s account of this project; it is not an independently established estimate of how long comparable verification would take in every case.
DeepMind also says it tested whether the system could identify insecure code by reintroducing two previously fixed bugs into an isolated version of Signal’s code. In both cases, it reports, the system produced a machine-checked proof that the altered version was insecure. The company says it will open-source the verification artifacts and has published a technical paper alongside its announcement.
From Protocol Design to Deployed Code
Formal verification can provide a stronger kind of assurance than testing alone: rather than checking selected inputs and outcomes, it uses mathematical proofs to establish that code meets specified properties. That distinction matters for cryptographic software, where a flaw in widely used code could put many users’ communications at risk.
DeepMind’s central claim is that the work connects security properties to the compiled implementation, rather than stopping at a protocol’s abstract design. In practice, code can fail to preserve a design’s protections through implementation errors. If the reported method can be independently examined and applied to other systems, it could make demanding verification work more practical for software used in communications, medical devices, or infrastructure. The announcement does not establish that those other applications have already been verified.
The work also illustrates a potential division of labor: AI agents generate and check large volumes of proof material, while human reviewers focus on specifications and areas flagged for attention. That may reduce some specialist workload, but it does not remove the need for experts to define the properties worth proving or assess whether the formal model matches the real system.
Products Worth Considering
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Anker SOLIX C300 DC Portable Power Station with 60W Solar Panel, 288Wh, 300W
As an affiliate, we earn on qualifying purchases.
Why Implementation Proofs Are Difficult
Cryptographic protocols underpin services including messaging, online banking, and web browsing. Researchers have long used formal methods to analyze security designs, but DeepMind says verification often stops at that level because proving that production code implements a design correctly can demand extensive manual work.
The Signal protocol is used to protect conversations in Signal and Google Messages. DeepMind says its researchers worked with feedback from Signal representatives on the methodology. The announcement describes the verified target as the protocol’s core cryptographic code and says the proofs cover properties including protection of messages before and after a device is temporarily compromised. It does not say that every component of either messaging app, or every possible security property, was proved.
DeepMind presents AlphaProof Code as an agent framework in which AI helps develop specifications and proofs, with Lean serving as the mechanical checker. Its claim of a first for verification of a widely deployed protocol is the company’s characterization of the project; the announcement does not provide a detailed comparison with all prior verification efforts.
“Formal verification—mathematically proving code runs correctly without flaws—is the highest standard of security assurance; however, it is difficult and often labor-intensive.”
— Google DeepMind
The Proofs’ Scope and Review
The announcement does not fully specify which files, dependencies, build configurations, or parts of Signal’s production software are included in the verification. It describes proofs of key security properties, not a guarantee that the entire application is free of vulnerabilities. A mathematical proof establishes that a formalized implementation meets a formalized specification; it does not by itself show that every real-world threat or implementation detail was represented in that specification.
It is also not yet clear from the announcement how independent researchers have evaluated the proofs, whether the reported four-week schedule includes all preparation and review, or how the claimed comparison with earlier verification work was assessed. The two bug tests are described by DeepMind, and details about the bugs and test procedure are not provided in the supplied announcement. The planned open-source release may help address some of these questions, but its contents and timing are not specified here.
External Review and Wider Testing
DeepMind says it is releasing the formal verification artifacts and points readers to a technical paper. Researchers and security specialists will be able to examine the specifications, proof files, and methodology as those materials become available. The company has not announced a schedule in the source material for outside evaluations or for a broader deployment of the framework.
The next test is whether the approach can be reproduced, whether reviewers agree that the specifications capture the intended security guarantees, and whether it can scale to other systems with different code and threat models. DeepMind says continued research is needed to scale the techniques and frames formal verification as a possible complement to automated defensive tools. No verification results for medical-device firmware, power-grid software, or other systems are reported in this announcement.
Key Questions
What is AlphaProof Code?
AlphaProof Code is Google DeepMind’s framework for using AI agents and human expertise to specify and prove software properties. DeepMind says its proofs are checked with Lean, a system for formal mathematical proofs.
What did DeepMind say it verified?
DeepMind says it proved specified security properties of the compiled code underlying the Signal protocol. The announcement does not claim that every component of Signal or Google Messages was verified or that all possible vulnerabilities were ruled out.
How long did the work take?
DeepMind reports that its verification process took four weeks. That is the company’s account of this project, not a general estimate for all formal verification work.
Does this prove Signal has no security flaws?
No. A proof can establish that modeled code meets specified properties. It does not establish that the entire app is free of flaws or that the specifications cover every possible threat. The announcement describes proofs of key properties, not a blanket security guarantee.
Can outside researchers inspect the work?
DeepMind says it will open-source the verification artifacts and has published a technical paper. The announcement does not specify when all artifacts will be available or describe completed independent reviews.
Source: deepmind.google
