DeepMind AI proves Signal's encryption secure in four weeks
What's the deal? Google DeepMind said it used a new AI system called AlphaProof Code to mathematically prove key security guarantees of the Signal protocol, the encryption behind billions of private messages in apps like Signal and Google Messages. AI agents powered by Gemini generated millions of lines of proofs in Lean, a language that mathematically verifies code. The four-week project found no new vulnerabilities in the parts it examined.
Why it matters: Formal verification — mathematically proving software behaves as intended — is considered the highest standard of security assurance, but it has long been slow and labour-intensive. DeepMind says this is the first time the implementation of a widely deployed protocol has been formally verified by any method, with or without AI.
How it works: Human experts worked with AI to translate Signal's security goals into formal specifications, reviewing every one. AI agents then produced the proofs and flagged areas needing a second human look — about 1% of the codebase. Lean automatically checked the rest.
What's the endgame? DeepMind pushed past the usual design-level checks to prove the compiled code running on devices meets properties such as keeping messages protected before and after a device is temporarily compromised. To test the system, researchers reintroduced two previously fixed bugs; in both cases it produced proof that the altered code was insecure.
What could go wrong? The work covers selected cryptographic guarantees, not the whole Signal app or every deployment condition, and humans still define and review what is proved. Research scientist Swarat Chaudhuri cautioned against a "vibe world where AI is generating the goals and AI is generating the proofs." For now, it remains a research project, not a commercial product.
The signal: The advance moves formal verification from checking a protocol's design to proving properties of its compiled code, potentially making rigorous verification far cheaper. It lands amid a broader push: Anthropic and startup Theorem recently reported AI-generated proofs across 123 open-source libraries, while firms including &TruthDealroom has a profile for this one. Try Dealroom → and Harmonic chase similar tools.
Read more: Ars Technica, Axios, deepmind.google
Image credit: Generated with Gemini
