Google DeepMind AI proves important Signal guarantees in 4 weeks
It found no new flaws, but it correctly found known flaws in Signal versions from before their repair.
Google DeepMind researchers used agentic AI, which works in steps without help, to prove with math that the Signal protocol keeps some important guarantees. The protocol protects messages in the Signal app, Google Messages and WhatsApp. The work was done in 4 weeks with a new framework, AlphaProof Code, and found no new flaws in the parts that it examined.
How it works
Agents write and correct proofs at almost each step. A person gives input and checks the work at important points. DeepMind had a head start from proof work that the Signal Foundation did before. The team releases the specification and the proof for use by other teams.
The proof
The analysis found no new flaws in the parts that it examined. On Signal versions from before the repair of known flaws, it found them and showed that the protocol could not be proved. Ehren Kret, chief technology officer of Signal, says the work gives more confidence in privacy.
What is not known
The scope was small. It did not prove the full Signal mobile app. Google declined to share the cost and declined to say if a tool will follow. Swarat Chaudhuri of DeepMind says there is much work is necessary before this works for all software.
Insight: DeepMind finds flaws and proves their absence
In November 2026, Google DeepMind said AI agents found 2 flaws in 1,024 bit RSA encryption. Its agents also prove Signal guarantees. This seems to put DeepMind on the 2 sides of security: tests that find flaws, and proofs that show none. Google declined to say if a tool will follow. Leaders: tell vendors to show which parts of their code have proofs.
Sources
Posted
