NederlandsGoogle DeepMind uses AI agents to verify Signal's encryption code
Google DeepMind has used AI agents to prove mathematically that the core of Signal's encryption meets key security requirements, right down to the code that actually runs on the processor. The work took four weeks and covered one-to-one chats, not the whole app.
Google DeepMind has used a new system, called AlphaProof Code, to check mathematically whether the core of Signal's encryption does what it should. According to Google's AI research lab, it is the first time that the working code of such a widely used encryption protocol has been proven at this level, with or without AI. The Signal protocol protects conversations in apps such as the Signal messaging app and Google Messages.
Two parts of the encryption
The team looked at two parts of libsignal, Signal's software library. The first part, PQXDH, sets up the secret keys at the start of a conversation, in a way that is also meant to resist future quantum computers. The second, the Triple Ratchet, creates new keys again and again with every message. This keeps messages protected, even if an attacker briefly had access to a device.
The researchers did not check the source code that programmers write, but the compiled program: the instructions that an ARM processor, the type of chip in most phones, actually carries out. This way, even an error that only appears when source code is turned into a working program is covered.
People decide what must be proven
Formal verification means using a mathematical proof to establish that software meets set requirements, instead of only testing it. First, the researchers used AI models to help translate Signal's security goals into precise mathematical requirements. Human experts checked every requirement.
Then AI agents running Gemini 4 Carbon, Google's latest top model, wrote the proofs in Lean: a language in which a computer checks every proof step by step. Teams of 40 to 80 agents each worked on one computer for one to three days at a time. Together they wrote about 6.5 million lines of proof. The agents flagged about 1 per cent of the work for an extra look by a person; Lean checked the rest automatically.
A team of five people did the work in four weeks. According to Google DeepMind, similar projects have traditionally taken expert teams years. The team built on earlier mathematical analyses of the protocol, including work that Signal itself had already commissioned. Signal's chief technology officer, Ehren Kret, calls it a major breakthrough that AI agents can turn working code into proofs that hold up.
No new weak spots came to light in the parts that were examined. To test whether the method really catches errors, the researchers put two old bugs, fixed long ago, back into a separate version of the code. In both cases the system proved that this version was not secure.
What is not yet covered
The proof assumes that the building blocks of the encryption itself, such as the mathematical key functions, work without errors. It covers conversations between two people, not group chats, and not the complete apps on phones. Group chats are on the list for later. Google DeepMind is making the requirements and proofs public, so that others can check them. Google would not say how much the research cost.
What does this mean for the road to the singularity?
AI agents have been writing code for some time, but here they proved that existing code works securely. Earlier this month, a thousand Claude agents showed that they can turn a famous mathematical proof into Lean; with Signal, the same kind of check is applied to software that millions of people use every day. If this can be scaled up, software that is proven to be reliable could also come within reach for things like medical devices and power grids, with people still deciding what has to be proven.
What is the technological singularity? Read the explanation →
