Formal verification is the least glamorous and most convincing form of software assurance: instead of testing a program on examples, you prove that it satisfies a specification, and a small trusted checker confirms the proof. Google DeepMind's new AlphaProof Code does this with a team of AI agents, and its first public target was Signal.
What happened. DeepMind introduced AlphaProof Code as a system of agents that produced machine-checked proofs in Lean about security properties of Signal's compiled libsignal code, the library that implements the protocol behind Signal's end-to-end encryption. The work took about four weeks. Comparable verification efforts have historically been measured in person-years.
The technical detail that matters is the word compiled. Most verification work reasons about source code. But the thing that runs on a phone is a binary produced by a compiler, and compiler optimizations can, in principle, introduce behavior the source did not describe. Proving properties about the compiled artifact closes that gap, and it is considerably harder.
How the trust works. The agents propose proof steps; Lean's kernel checks them. A hallucinated lemma simply fails to check, so the system cannot fake a proof the way a language model can fake a citation. The trust anchor is the checker, not the model, which is why this class of result is more durable than a benchmark score. The weak point moves somewhere else: the specification. A verified proof that the code does what the spec says is only as valuable as the spec's description of what security actually requires.
What it does not prove. A passing proof does not mean libsignal is exploit-free. Side-channel behavior, timing, memory safety at the boundaries, and the protocol logic that lives above the library are separate questions, and each needs its own specification. The result also says nothing about the client applications millions of people actually use, which are the more common attack surface.
The timing is pointed. Earlier this month OpenAI's restricted GPT-6 Cyber model was reported to have found 948 vulnerabilities in the Linux kernel and dozens of previously unknown high-severity issues in Firefox, results the company reported itself. Offense and defense are being automated in the same season, by the same class of technique: search a large space of possibilities with a model that can check its own work.
That symmetry is the interesting part. Vulnerability hunting gets a model that reads code and proposes inputs. Verification gets a model that reads code and proposes proofs. Both scale with compute, both depend on a checker to filter nonsense, and both change the economics of work that used to be rationed by the scarcity of experts.
What to watch. Whether other critical infrastructure, TLS stacks, kernels, password managers, gets the same treatment, and at what price per proof. Whether labs publish their specifications alongside their results, since a proof against a secret spec is a press release. And whether the security industry starts treating formal verification as a procurement requirement rather than a research curiosity. If a four-week proof becomes a line item, the vendors selling "we take security seriously" will need a new sentence.
DeepMind's AlphaProof Code Formally Verified Signal's Encryption in About Four Weeks
← Previous: AI Reads Gödel's Shorthand Notebooks and Surfaces Theorems Nobody Had Catalogued
Next: →Google Opens Pre-Orders for $395 Gemini Glasses, Shipping December 10 →