Google DeepMind used AI agents to formally verify key guarantees of the Signal protocol in four weeks
Its new AlphaProof Code framework found no new flaws, correctly failed older Signal versions with known bugs, and the spec and proof are being released.
DevIsDead Newsroom
Google DeepMind researchers have used agentic AI to mathematically prove a number of the core cryptographic guarantees in the Signal protocol, Lily Hay Newman reports for WIRED. The project took four weeks and used a new framework called AlphaProof Code.
Signal's protocol protects messages in the Signal app, Google Messages and WhatsApp, so an independent proof matters well beyond one app.
What formal verification buys you
Formal verification uses math to show code does what it's supposed to, with no critical holes. It has long been a security ideal and a slog: people had to describe the properties they wanted, turn them into formulas, then build the proof by hand. Agents speed up that last step a lot.
What they found
No new vulnerabilities in the parts they examined. As a sanity check, the team pointed AlphaProof Code at older Signal versions with known, long-fixed bugs. It caught them and showed those versions could not be verified. DeepMind is publishing the specification and proof so others can build on them.
The scope was limited. It didn't cover the whole Signal mobile app, and the team had a head start from earlier verification work by the Signal Foundation. Signal CTO Ehren Kret said DeepMind's approach expands on earlier work that covered smaller pieces of the code.
Humans stay in the loop
Research scientist Swarat Chaudhuri described the method as human-AI collaboration: agents work and correct themselves through most steps but ask people for input and review at key points. He doesn't want a setup where AI writes both the goals and the proofs, like a student grading their own test.
The cost question
Whether this scales, and at what price, is open. DeepMind did the Signal work for free as research, Google wouldn't share what it cost, and it wouldn't say whether a commercial product is planned.
Scoreboard update: AI +1. AI agents prove a widely used encryption protocol's key guarantees in four weeks, with humans reviewing.
Stay undead.
Comments 0