- Google DeepMind says AI agents produced a machine-checked proof of a number of key guarantees in parts of the Signal protocol in four weeks, finding no new flaws in the parts examined.
- The result shows that when AI generates the work, assurance depends on a trusted checker and on human review of the goals being proved.
- Ask vendors who wrote the specification, what a proof leaves out, and which tool checks it.
The proof is not the part you have to trust
Security assurance usually rests on people. An expert reads the design, thinks hard and signs off. That is slow, and it does not scale to the amount of code now being written, much of it with AI help.
A project from Google DeepMind points to another arrangement. As reported by security researcher Bruce Schneier, DeepMind used AI agents to formally verify a number of the Signal protocol's key cryptographic guarantees. Formal verification means proving mathematically that a design meets a stated goal, rather than testing it and hoping.
The lesson here is that an AI can generate the work while something else vouches for it. The trust moves away from the AI and toward the checker and the question being asked.
What DeepMind says it did
The account comes from DeepMind's own description, quoted by Schneier. It describes a proof-of-concept completed in four weeks, using a new framework called AlphaProof Code.
The work had three steps. First, researchers worked with a mix of models to turn Signal's security goals into formal mathematical specifications. Human experts reviewed every one of them.
Second, agents powered by Gemini Carbon wrote the proofs. DeepMind says they produced over six million lines of Lean. Lean is a language in which a program can check a proof mechanically, line by line.
Third, the agents audited their own proof and flagged the areas that needed a second look from a person. DeepMind says that was just one percent of the codebase. Because Lean checks the rest automatically, the experts could focus on that one percent.
Why a machine-checked proof changes the review
One reader on Schneier's post asked who could possibly read six million lines. The answer is that nobody has to read them all. A person does not need to trust the AI that wrote a proof, because Lean accepts or rejects every step.
That makes the AI's output cheap to produce and cheap to doubt. A wrong proof fails the check. A correct one passes, however it was made.
This holds only if the goal is stated correctly. A proof shows that a design meets the specification. It cannot show that the specification captures what you care about. DeepMind's account puts human experts at exactly that point, reviewing every specification.
What the result does and does not show
DeepMind did not find new vulnerabilities in the portions of Signal it examined. The write-up describes this as further independent validation of the cryptography behind Signal, Google Messages and WhatsApp.
Two limits apply. The work covered only parts of the protocol and a number of its guarantees, not the whole system. And this is a single proof-of-concept, reported by the team that built the tool.
One result speaks to the method itself. When researchers ran AlphaProof Code on older Signal versions with known, long-fixed flaws, it correctly identified them. It also showed that those versions could not be formally verified. A verifier that fails on known-bad designs is more credible than one that has only ever passed good ones. The write-up does not say how many older versions were tried.
DeepMind is also releasing the specification and proof for others to use. That lets outsiders inspect the goals as well as the conclusion.
Questions to put to your team and vendors
Most organisations will not verify a protocol themselves. They will buy products whose makers claim some form of AI-assisted assurance. Three questions follow.
Who wrote the specification, and who reviewed it? A proof of the wrong goal is worth little.
What does the proof leave out? Ask for the boundary in writing: which components, which guarantees, which versions.
Which tool checks the result, and can someone independent run it? Assurance that rests on the vendor's own AI is a claim. Assurance that rests on a published specification and a mechanical checker is something a third party can test.
AI will write more of the evidence that software is safe. The useful habit is to stop asking whether the AI is trustworthy, and to ask what is checking it.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Schneier on Security.





