News · Research and benchmarks · News
Google DeepMind AlphaProof Code: Signal crypto verified in 4 weeks
Google DeepMind announced AlphaProof Code this week. It is an agentic framework for formal verification, the practice of mathematically proving that code runs correctly. Researchers used it to verify the cryptographic core of Signal, the protocol that protects messages in apps like Signal and Google Messages.
The work is notable because it went from design level to implementation. Agents generated over six million lines of Lean, a proof language, in four weeks. Human experts reviewed the specifications and about one percent of flagged code.
In one line: Google DeepMind's AlphaProof Code used AI agents to generate six million lines of Lean proofs and verify Signal's cryptographic core in four weeks, a task that usually takes years.
What AlphaProof Code is and how it worked
AlphaProof Code is a multi-LLM agentic framework. It splits formal verification the way a research team would. Human experts work with models to translate security goals into formal specifications. Then agents powered by Gemini Carbon do the proofs. The agents wrote over six million lines of Lean. Lean automatically checks the rest, so humans only needed to review about one percent of the codebase. This is the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth, according to Google DeepMind. It is also the first time the implementation of a widely deployed protocol like this has been formally verified by any method.
For teams that run agentic tools, the process is closer to the Gemini 4 Carbon release than to a single model call. I keep a list of similar agent frameworks in the AI agents directory on this site.
Why the compiled implementation matters
Most protocol verification stops at the design level. AlphaProof Code went further. It proved that the compiled code fulfils the key security properties, such as keeping messages protected before and after a device is temporarily compromised, down to the exact code running on devices. Real world vulnerabilities often hide in the gap between design and implementation. Bridging this gap makes the proofs much more complex. Google DeepMind says the verification process took just four weeks, a task that historically takes expert teams years.
This matters to anyone who ships software. The hard part has always been proving that production code matches the cryptographic design, not just talking about the design. AlphaProof Code automatically translated live code into valid Lean theorems, which Google DeepMind calls a breakthrough.
Bug checks and what they found
Google DeepMind reintroduced two real, previously fixed bugs into an isolated version of Signal's code. In both cases, AlphaProof Code produced a machine checked proof that the buggy version was insecure. It did not find any new vulnerabilities in the portions of Signal it looked at. This is a useful sanity check: the system can tell correct code from buggy code.
The researchers are open sourcing the verification artifacts for community use. Signal's CTO Ehren Kret says rigorous verification gives greater confidence in privacy guarantees as AI driven security threats grow.
What this does not do yet
This is a milestone, not a full product. The verification did not cover the entire Signal mobile app. DeepMind researchers tell Wired that there is a long distance to go before they can scale this to all software. The project also had a head start from prior formal verification work by the Signal Foundation. Human experts still need to define what to prove. You cannot point AlphaProof Code at an arbitrary codebase today and get a security certificate.
For my client work, where I ship RAG chatbots and internal AI tools, nothing changes immediately. My code is not a cryptographic protocol. But the direction is real. If formal verification of critical code gets cheaper, regulated clients may start asking for similar proof on the cryptographic libraries they use.
What I would try next
I would watch two things. First, the open source verification artifacts. If they include the Lean specifications around device compromise, that is a good learning resource for teams that want to write machine checkable security requirements. Second, the pricing and access path for agents built on Gemini Carbon. If the cost per proof drops, formal verification could become part of a security review checklist for fintech or health clients.
The honest answer for this week is: nothing to do yet, just watch. I would not rewrite any delivery plan around this. But I would add AlphaProof Code to the list of things that can turn an AI agent swarm into a checking process, not just a generation process. For a similar structure around agentic coding delivery, I wrote about shipping client products without a dev team.
Quick answers
What is Google DeepMind AlphaProof Code?
Google DeepMind AlphaProof Code is a multi-LLM agentic framework for formal verification. It combines human experts with AI agents to write mathematical proofs in Lean and verified Signal's cryptographic core.
Did AlphaProof Code find new vulnerabilities in Signal?
No, AlphaProof Code did not find new vulnerabilities in the portions of Signal it verified. When researchers reintroduced two known bugs, the system correctly proved the buggy code was insecure.
How long did the Signal verification take?
The verification process took four weeks, according to Google DeepMind. The company says this task historically takes expert teams years.