Monday, November 16, 2026

Top 5 This Week

Related Posts

AlphaProof Code: DeepMind’s AI mathematically proves the security of Signal’s encryption


Saying that software is secure is one thing; demonstrating it mathematically is another, much more demanding task. This is what we have just realized Google DeepMind on one of the most widely used encryption software in the world. The laboratory presented the November 10, 2026 AlphaProofCodea system that combines human expertise and AI agents, and applied it to the cryptographic core of Signalend-to-end encrypted messaging. Result: key safety properties proven in four weekswork that DeepMind says historically takes teams of experts years to complete.

What is formal verification?

Most software is tested: we run it on numerous cases and check that it behaves well. But a test only covers the situations envisaged. There formal verification goes further: it consists of demonstrating, like a mathematical theorem, that a program respects a property in all possible casesaccording to an attacker model defined in advance.

This method is renowned for its reliability, but also for its cost. Writing these proofs by hand takes so long that it is usually reserved for critical systems, such as aeronautics. It is precisely this bottleneck that DeepMind wants to break with AI.

A distribution of roles between humans and AI

AlphaProof Code is described by DeepMind as a framework bringing together multiple language models into an agent system. The division of work is clear:

Actor Role
Human experts Decide what to prove and write formal specifications, all proofread by humans
AI Agents Write detailed evidence and identify passages that require human review
Lean Mechanically verifies each proof, without letting any errors slip through

According to DeepMind, the agents only reportedabout 1% of the code for human examination. The evidence produced exceeds six million linesa volume that no human team could write in a few weeks.

What has been proven on Signal

The target is libsignalthe reference library that implements the Signal encryption protocol. Notably, the verification does not relate to an abstract description of the protocol, but to the compiled machine codethat is to say the instructions actually executed by a 64-bit Arm processor, that of modern smartphones. DeepMind has therefore modeled the processor that executes this code in Lean.

The properties demonstrated concern in particular the fact of keep messages protected before and after a device is temporarily compromised. According to the specialist site CyberInsider, the evidence covers the phase of establishing the connection between two devices and the mechanism which constantly renews the encryption keys, including in a scenario where a future quantum computer breaks part of classical cryptography.

To test the method, the researchers reintroduced two real bugs, fixed sincein an isolated copy of the Signal code. In both cases, the system produced machine-verified proof that the modified version wasn’t sure. No new vulnerabilities have been reported in the examined parts.

Why check machine code rather than source code?

Developers write their programs in a human-readable language, here Rust. Software called a compiler then translates this source code into instructions for the processor. However, this translation is not neutral: optimizations, choice of the compiler or subtleties of the language can produce behavior slightly different from that which we believe to have written. Proving the security of source code therefore leaves a gray area between what is written and what actually executes.

By working directly on machine code, DeepMind reduces this gap: the proof relates to the instructions that the processor actually executes. This is one of the points that distinguishes this work from the verifications carried out so far on abstract models of the protocol.

What the evidence does not guarantee

A formal proof is only valid within the framework defined at the start. Several limits are specified by DeepMind and in the published files:

  • the verification concerns the libsignal library compiled in a command line test environment, and no on Signal applications as distributed;
  • THE cryptographic primitives basic, like the encryption algorithms themselves, are modeled and considered reliable, without being verified in turn;
  • attacks by auxiliary channelswhich exploit, for example, calculation time or electricity consumption, are not covered, according to CyberInsider;
  • guarantees depend on attacker models retained: a so-called symbolic model and a computational model, closer to real cryptography.

In other words, the proof does not make Signal invulnerable: it establishes that, in this specific context, the code behaves as expected.

Results verifiable by everyone

DeepMind published all the evidence in open source under Apache 2.0 license, on its GitHub platform. Anyone with Lean can recheck them, provided they have patience and a powerful machine: compilation and verification take several hours and use up to several tens of gigabytes of memory. In contrast, the AlphaProof Code framework itself has not been announced as available to the public.

Signal was involved in the process. Three of its engineers are thanked for their feedback on the method, and its technical director, Ehren Kret, cited by DeepMind, welcomes “a major breakthrough”: using AI agents to translate code in production into valid Lean theorems. According to him, such rigorous verification gives even greater confidence in the confidentiality guarantees of the code.

When AI is used to prove rather than guess

This work is part of an underlying trend: using AI not only to produce content, but to demonstrate machine-verifiable results. Mathematics has already benefited, such as with the KLS conjecture solved with the help of AI in October. In cybersecurity, AI is also used to find vulnerabilities, like the 948 vulnerabilities in the Linux kernel exploited by OpenAI agents. Formal verification takes the problem from the other end: proving the absence of certain flaws rather than looking for them.

DeepMind believes that this work lays the foundation for large-scale verification of critical software, such as medical device firmware or software that controls power grids. If the method becomes widespread, the question will no longer only be whether software has been well tested, but whether it has been proven. It remains to be demonstrated that it applies with the same effectiveness to systems less well documented than Signal, whose protocol has been studied by cryptographers for years.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles