All news

Google DeepMind formally proves the security of the Signal cryptographic protocol using SI agents

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

What expert teams spent years on, SI agents did in a month: mathematically proven code security stops being a luxury for a select few systems and can become standard practice for banks, medicine and energy.

Google DeepMind has presented the AlphaProof Code framework, in which people together with models formulate exactly what needs to be proved, and then a group of SI agents based on Gemini Carbon writes formal proofs in the Lean language. With it, researchers proved in four weeks the key security properties of the code that the Signal protocol runs on (it protects conversations in Signal and Google Messages). According to DeepMind, this is the first time the implementation itself of a widely used protocol has been formally verified, rather than its design — and independently of whether SI was used. The agents generated more than six million lines of Lean; automated checking left about 1% of the code for humans to review. Testing against two real, already-patched vulnerabilities reintroduced into an isolated version of the Signal code showed their insecurity to be unprovable. The artifacts have been released publicly. Limitations: it is the protocol core that was verified, not the entire Signal mobile app; part of the initial work had already been done by the Signal Foundation; Google is not disclosing the project's cost and is not announcing a commercial product. "There's still a long way to go before scaling this to all the software in the world," said DeepMind researcher Swarat Chaudhuri.

History

Google's multi-agent research and formal proofs

2 events
  1. Google shows agent teams solving open scientific problems with /teamwork
  2. Google DeepMind formally proves the security of the Signal cryptographic protocol using SI agentsOpen now