Hopp til hovedinnhold

Google DeepMind had AI agents prove Signal's encryption holds – all the way down to machine code

AlphaProof Code wrote 6.5 million lines of machine-checked proof for the Signal protocol in four weeks. The proof covers the core code, not the whole Signal app.

Av Joachim HøgbyKilde: Google DeepMind

Google DeepMind has used AI agents to mathematically prove that key security properties of the Signal protocol hold, all the way down to the machine code that actually runs. The protocol protects messages in Signal and Google Messages. According to DeepMind, it is the first time the code behind such a widely deployed encryption protocol has been formally verified at this depth, with or without AI.

The work took four weeks. DeepMind says it would previously have taken expert teams years. The company published the result and a technical report on 10 November, and highlighted it on its official X account on Wednesday 11 November at 11:50 UTC.

How DeepMind did it

The tool is called AlphaProof Code. It splits the work the way a research team would. First, the researchers, working with several language models, wrote Signal's security goals as formal mathematical specifications in the proof language Lean. Humans reviewed every specification.

Agents powered by Google's Gemini Carbon model then took over. They wrote about 6.5 million lines of Lean proof, according to the technical report. A proof is accepted only if Lean's own kernel checks it against the approved specification. The agents themselves flagged about one percent of the material for an additional human review.

The proofs rest on roughly 26,000 lines of specifications and a formal model of ARM machine code (AArch64), which the researchers have to trust. They do not have to trust the Rust source code or the compiler, because the proof covers the finished program.

What has actually been proven

The proofs cover Signal's post-quantum handshake PQXDH and the Triple Ratchet messaging mechanism in libsignal version 0.105.4 with SPQR 1.6.0. Among other things, DeepMind proves authentication, that past messages stay protected if a key leaks later, and that a conversation regains its security after a device has been compromised for a period.

To test that the tool can tell good code from bad, the researchers reintroduced two real bugs that Signal had already fixed. In both cases the system produced a machine-checked proof that the buggy version was insecure.

In DeepMind's announcement, Signal's chief technology officer Ehren Kret calls using AI agents to translate live code into valid Lean theorems "a major breakthrough".

What the proof does not say

A headline saying "Signal is proven secure" goes too far. The verification covers a narrowly scoped command-line build of the library, compiled for AArch64. Key generation is not verified, and the cryptographic building blocks are described as contracts that are assumed to hold. The Signal app as a whole, its servers and group chats are not part of the proof.

DeepMind does not disclose what the work cost in compute, and does not say whether AlphaProof Code will become a commercial product. The company is making the specifications and proofs openly available.

What business leaders should know

Formal verification has been reserved for aviation, nuclear power and a handful of critical systems, because it has been too expensive. If AI agents can do the job in weeks instead of years, mathematical guarantees for code could become something more organisations ask for. That is our assessment, not a promise from DeepMind.

  • Encrypted messaging now has a stronger proof base. For organisations that use Signal for sensitive communication, this strengthens confidence in the core of the protocol, but not in the phone, the app or the routines around it.
  • Ask vendors how their code is checked. DeepMind itself describes the goal as bringing this level of assurance to businesses, governments and critical infrastructure. It becomes a natural question when buying critical software.
  • The same tools cut both ways. Earlier in November, hogby.ai reported how DeepMind's agents found weaknesses in legacy RSA-1024. AI is becoming both an attack tool and a defence tool for cryptography.

Sources and media

Illustration: AI-generated with Grok Bot / hogby.ai

Mer om AI

Oppskriften. Strategien. Hver dag.

AI-nyheter for ledere. Kuratert av en CIO som bygger det selv. Daglig i innboksen.