What DeepMind actually showed
On 10 November 2026, Google DeepMind announced AlphaProof Code, an agentic framework that combines human experts with AI agents for formal verification. Its showcase was the core of Signal's cryptographic code. According to DeepMind's blog, agents powered by Gemini generated over six million lines of Lean proof. Lean then machine-checked the compiled code against key security properties, including protection of messages before and after a device is temporarily compromised. The work took four weeks. The Next Web's report, which cites Wired, adds that the proofs cover forward secrecy, authentication and post-compromise security in one-to-one chats.
The team also tested whether the method could catch real flaws. They reintroduced two previously fixed bugs into an isolated copy of the code. DeepMind says that in both cases the system produced a machine-checked proof that the buggy version was insecure. That was a validation test: the method caught both reintroduced bugs. It found no new flaws in the portions it checked.
The limits matter. The proofs cover the researchers' own build of the library, not the Signal apps. Group messaging and key generation were not verified, and the proofs assume the underlying cryptography is sound. The proofs rest on human-reviewed specifications and on a model of the processor, so a mistaken specification would not be caught by the proofs. DeepMind's own researcher, Swarat Chaudhuri, told Wired there is "a long distance to go before we are able to scale this to all the software of the world." This is a strong result about one protocol. It is not proof that every messaging product is secure.
The case for lawful access, stated fairly
Law enforcement's concern is real and measurable. The Congressional Research Service's lawful access report (IF11769), updated 29 October 2026, cites the Administrative Office of the U.S. Courts' 2025 Wiretap Report. Judges authorized 1,735 wiretaps in 2025, and investigators encountered encryption in 525 of them. They could not decrypt 486, roughly 93%, up from 88% in 2024. Officials call this "warrant-proof" encryption. They argue that a court-authorized search should not be defeated by design, and that a society which lets judges authorize wiretaps should not accept a class of communications that is unreachable in principle.
That argument deserves a real answer, not a slogan.
Why verification changes the terms of the debate
For years the lawful-access debate has rested on a prediction. Engineers said no one can build a door that only the right people can open, and officials replied that smart engineers could. The same CRS report notes that "researchers have yet to demonstrate how it would be possible to create a door that could be accessed only in lawful circumstances." Until now, critics could still say that the existing designs themselves were unproven.
DeepMind's result narrows that gap. When the baseline is a design whose key properties are machine-checked, the policy question is no longer "is encryption trustworthy?" It becomes "what does mandated access cost relative to a verified baseline?" A mandate would force a different design. It would have to break exactly the properties the proofs establish, such as post-compromise security, which protects users after an attacker briefly gains access to a device. A mandated mechanism would not be a neutral feature added to a sound system. It would be a chosen weakening, and the proofs would no longer apply to the result.
The cost does not fall only on suspects. A mechanism that gives lawful actors a path in is also a path attackers will look for. The EFF has long documented this in its Crypto Wars materials, recalling the 1990s fight over the Clipper Chip. A verified design also shows what a successor would be measured against. If a regulator wanted an exceptional-access scheme, the burden would fall on the proponents to show it could meet the same standard of proof. No one has done that.
What this does not settle
Two cautions keep the argument honest. First, the proofs concern the protocol and its implementation. They do not cover endpoints, such as a phone with spyware, or metadata, or cloud backups. Those are the places where proportionate, targeted tools already operate, so verification does not make investigators blind. It narrows the debate to what a government can lawfully do at the edges of a sound system. Second, I did not find any direct U.S. government response to the DeepMind announcement, and none should be inferred. The result is evidence for the policy debate, not a ruling in it.
A proportionate way forward
A pro-innovation policy does not demand that engineers dismantle a verified design to serve an investigative preference. Several steps are more proportionate:
- Invest in lawful capability that does not break the protocol. That means device-level forensics, targeted court-authorized hacking with warrants and oversight, and better handling of the metadata that remains available.
- Make exceptional-access proposals prove their security. Any mandate should have to meet the same formal standard that Signal's core code now meets, and publish specifications and proofs for review.
- Use real data. The 93% undecryptable figure should be paired with outcome data. Courts and agencies should report how often decryption failure changed a case, not just how often it occurred.
- Protect verification itself. Open-sourced proofs and specifications, like DeepMind's, let independent researchers check the work. Policy should not discourage this scrutiny.
The lesson of the DeepMind result is modest but useful. Strong encryption is no longer just a design that has survived years without a public break. For its core protocol, it is increasingly something that can be checked. Mandating a hole in a design that can be shown to be sound is a choice, and lawmakers should weigh it as one.