17 November 2026Published by Silicon Valley InnovationOwnership & standardsStandards
Technology news from BangkokBANGKOK / TH + EN
SV News
ไทย

Security

DeepMind's AI 'proved' Signal's encryption in 4 weeks

The proof runs to more than 6 million lines and reaches the code the chip actually runs, but it does not cover the whole app, group chats or a stolen phone. Signal users can check for themselves that no one has slipped into a chat.

Sources

อ่านภาษาไทย

A black smartphone screen with Signal's white speech bubble logo in the centre, held in one hand, with a blurred living room and grey sofa behind
A phone showing the Signal app; illustrativePhoto by Mika Baumeister on Unsplash
Jump to a section
  1. Testing finds only the problems you look for, proof covers every case
  2. AI wrote more than 6 million lines of proof in 4 weeks, people re-checked about 1%
  3. The proof does not yet cover the whole app or your phone
  4. What chat users can do today
  5. FAQ
  6. Sources

Every time a chat app says a conversation is encrypted, we can only take its word for it, because no one is going to read hundreds of thousands of lines of code themselves. Even the companies that write the apps can only test them again and again and hope nothing slips through.

On 10 November 2026 Google DeepMind launched AlphaProof Code, a system in which AI agents, AI that can work through many steps on its own, helped prove with maths that the encryption code behind Signal really meets key security properties. The proof goes down to the compiled code, the instructions the chip in a device actually runs, rather than stopping at a design on paper, according to DeepMind's blog.

Testing finds only the problems you look for, proof covers every case

The word "proved" in this story carries far more weight than "tested".

Think of the padlock on your front gate. Testing means trying 1,000 wrong keys one by one. If none opens it, all you know is that those 1,000 keys do not work; you still do not know about key 1,001. Proof means going through the mechanism inside the lock piece by piece until you can show that no key except the right one will ever open it, however many millions of keys anyone tries.

The tool that checks the proof is Lean, a computer language for writing mathematical proofs. It works like a teacher who marks every line of working instead of just the final answer. If any line skips a step or is slightly wrong, Lean rejects the whole proof. A proof that passes Lean therefore does not rely on trusting whoever wrote it: a person and an AI are checked by the same standard.

Several properties were proved, according to DeepMind's project page. The easiest for chat users to grasp is forward secrecy: if your keys leak one day, old messages you have already sent still cannot be read. Another is post-compromise security: if a device was briefly hacked, new messages become safe again once the intruder is gone.

The others are confirming that the person you are talking to really is that person, and the secrecy of the handshake, the moment two devices first agree on a secret key before a chat begins. All of it was proved on libsignal, the code library that encrypts chats between two people. DeepMind wrote that the Signal protocol protects billions of private conversations in apps such as Signal and Google Messages.

AI wrote more than 6 million lines of proof in 4 weeks, people re-checked about 1%

What got this work noticed is its size and speed.

Proving code like this has taken teams of experts years. This time it took 4 weeks, with the work clearly split. People decided what to prove and checked every specification written in Lean. AI agents running on the Gemini Carbon model wrote the proofs, producing more than 6 million lines. The project page gives a more precise figure of about 6.5 million lines.

How much is 6 million lines? Reading one line a second without sleeping, one person would need about 70 days to finish. No one can check that by eye, so the checking falls to Lean, which does it automatically, while the AI points out places that deserve a second human look. Those came to about 1% of the code.

The team also tested whether the system really catches faults. They put 2 old bugs, which Signal fixed long ago, back into a separate copy of the code. Both times the system proved that the buggy version was insecure. It is like striking a match under a smoke alarm to see that it really sounds, rather than one that says everything is safe.

On Signal's side, Ehren Kret, Signal's chief technology officer, called the use of AI agents to turn live code into theorems that can be checked a major breakthrough, because earlier proof work covered only small parts of the code. Signal's team also gave feedback on the method, but the proof work was DeepMind's.

The proof does not yet cover the whole app or your phone

A proof is only as strong as the scope it was given.

What was proved is libsignal version 0.105.4 wrapped in a test program, not the whole Signal app on a phone, as Let's Data Science summarised from WIRED's reporting. DeepMind's own project page lists what is out of scope: key generation, group chats, using one account on several devices, and side channels, which means guessing secrets from side traces such as how long a device takes to compute or data left behind in memory.

Some basic encryption building blocks were also assumed to work correctly rather than proved again. Google did not disclose what the work cost, and Swarat Chaudhuri, a DeepMind researcher, acknowledged there is still a long way to go before the method can be applied to other software at a workable cost.

For chat users, the point to remember is that this proof is about messages travelling from one device to another. If someone gets hold of your unlocked phone, or the phone has spyware on it, they can read the chats on screen just as you do, and the best encryption cannot help with that. Other apps that use the same protocol also have code of their own that this proof does not touch.

What chat users can do today

This work does not change how you use Signal, but it is a reminder that the other half of security sits on your own device.

Start by updating every chat app to its latest version, because fixes like the 2 bugs the team put back for the test only reach your phone through updates. Then set a screen lock passcode on your phone, because no proof can protect a phone anyone can pick up and open.

Signal users can check for themselves that no one has slipped into the middle of a chat. According to Signal's blog from August 2026, open the profile of the person you are talking to, tap View Safety Number, then tap Verify automatically under the Automatic key verification heading. If it passes, you will see a green check mark and the words Encryption verified. If the automatic check is not available, for example because you do not have the other person's phone number, compare safety numbers the old way, in person or through another channel you trust.

This week, 17 to 23 November, open the chat with the person you discuss important things with most often and check that chat's safety number. It does not take long, and it tells you more than a line saying the chat is encrypted.

FAQ

Did the AI prove that Signal has no bugs at all?

No. The proof covers some security properties of libsignal inside a test program. It does not cover the whole app, group chats, key generation or using several devices.

Do people who use WhatsApp or other apps benefit too?

Some other apps use the same protocol as Signal, but this proof was done only on libsignal's code. The code each app writes for itself was not proved along with it.

If my phone is stolen, does encryption still help?

Encryption protects messages in transit. If someone else gets your unlocked phone, they can read the chats on screen just as the owner can, so set a screen lock passcode and keep the phone updated to the latest version.

Sources

People, companies & topics

Signalencryptiononline safetyAI agents

More from SV News

Share to Stories

Save this image or choose an app on your phone. Add the article URL with Instagram's Link sticker.

Preparing image...

Card colour
Card photo
Card text (choose up to 3)

Search

Saved

Saved in this browser

Your reading list starts here. Save a story to come back to it later.

Story photography