Skip to content
MaungaAI news, explained

Safety and security

Signal passes an AI-run math exam, but only in its key parts

AI agents from Google DeepMind took four weeks to mathematically prove several security guarantees of Signal's encryption. They did not review the whole app, and no one knows yet how much it would cost to do this at scale.

2 min read3 sourcesTranslated with AI from the Spanish original

Share: WhatsApp · Email

AI-generated illustration for Maunga. Not a photograph.

The essentials

  • Claimed by one partyAccording to DeepMind, its AI agents took four weeks to mathematically prove several key guarantees of Signal's encryption protocol (WIRED, Xataka).
  • Claimed by one partyThe analysis found no new flaws in the parts it reviewed. When given old versions with flaws that had already been fixed, it did catch them (WIRED, Xataka).
  • The full mobile app was not verified, and earlier verification work by the Signal Foundation gave DeepMind a head start (WIRED, Xataka).
  • Claimed by one partyAccording to DeepMind, the agents checked their own work and flagged about 1% of it for a person to review (Xataka).

Why it matters

Formal verification means using math, rather than one-off tests, to check that a program does what it should. It is considered the strongest guarantee, but it used to take years of expert work (Xataka). If AI cuts that to weeks, banks and critical services could start asking for guarantees like this, as DeepMind suggests (WIRED). The researchers themselves admit it is still a long way from being applied to all software (WIRED).

What does it mean for me?

If you use Signal, you don't need to do anything. It's one more check that the encryption protecting your chats does what it promises, at least in the parts that were reviewed.

The details

The system is called AlphaProof Code, and it combines several AI models with human experts (Xataka). Researchers and models translated Signal's security goals into precise mathematical rules, and a person reviewed each one. The agents then wrote the proofs in Lean, a language that lets a computer check them on its own (Xataka). DeepMind says this is the first time the real code of such a widely used protocol has been verified this way (Xataka).

The team wanted to see whether the system could tell good code from bad. To do that, it put two real, already fixed flaws back into an isolated copy of Signal, and in both cases the system proved that version was not secure (Xataka). DeepMind did the analysis for free, as a research project. Google has not said how much it cost or whether it will sell this service (WIRED).

What we don't know

  • How much the analysis cost, and whether it would be affordable for much larger programs.
  • Whether Google plans to sell formal verification as a product or service.
  • Whether experts outside DeepMind have already reviewed the proofs it has published.

Sources

This story is based on 3 sources. In the text, each fact names in brackets who reports it.

Was this story useful? Yes Not really

Spanish original