Code & dev Story
Google DeepMind AI agents produce formal proofs of security guarantees in Signal's libsignal library
Google DeepMind's AI agents, called AlphaProof Code, produced formal proofs of the security guarantees of libsignal, Signal's encryption library, over four weeks with teams of up to 80 agents, according to Numerama (10 November 2026).
This story was written with the help of artificial intelligence, checked automatically, then reviewed by an editor before publication, under our AI policy.
What DeepMind did on libsignal
Google DeepMind AI agents, called AlphaProof Code, produced formal proofs of the security guarantees of libsignal, according to Numerama. Libsignal is the library that implements Signal's end-to-end encryption protocol. Variants of that protocol also power WhatsApp, Messenger and Google Messages.
The work lasted four weeks. Teams of up to 80 agents ran on the machines of five Google DeepMind researchers. The report does not say more about how the agents were organised or how the work was divided among the five researchers.
Proofs checked by machine
The agents, which use the Gemini 4 Carbon model, wrote about 6.5 million lines of proof in the Lean language. Lean's checker verifies every step automatically, so trust rests on that check rather than on the reliability of the agents themselves.
According to the technical paper, the proofs establish that the compiled libsignal program implements the key-renewal mechanism (the ratchet) and provides its protections. These include protection against theft of keys stored on a phone being used to read old messages. This is the authors' own description of their results, and it has not been independently assessed in the material available.
A test with old bugs
To validate the method, the researchers reintroduced two old libsignal bugs, since fixed. The system showed that those versions no longer offered the expected guarantees. The method therefore detected known flaws, which says nothing about flaws not yet known.
The test shows that the proofs can fail when a guarantee is broken, which is a necessary property for a verification method. It does not, on its own, measure how many real-world flaws such a method could uncover.
The limits of the scope
The verification covers only two-party conversations. It assumes that the basic cryptographic components work, without verifying them. It also applies to a Linux test build, not to the mobile apps people use every day.
A formal proof therefore guarantees a specific property of a specific object. It is not a global security certificate for Signal or for the messaging apps derived from its protocol.
Because variants of the protocol also run in WhatsApp, Messenger and Google Messages, the subject extends beyond Signal. Still, the reported scope covers libsignal itself, and nothing in the available material says that the proofs apply to those other products.
Reactions and open points
Signal's chief technology officer, Ehren Kret, welcomed the work as "une avancée majeure" (a major advance). The authors, for their part, conclude that as proofs become cheap, "la question difficile devient de savoir quoi prouver" (the hard question becomes what to prove).
One point remains to be documented: the official DeepMind page that was consulted contains no detailed article on AlphaProof Code. It only lists other posts (CodeMender, Gemini 4 Carbon, and others). The technical details therefore rest, for now, on Numerama's account.
FAQ
What is libsignal?
It is the library that implements Signal's end-to-end encryption protocol, variants of which also power WhatsApp, Messenger and Google Messages.
Are the Signal mobile apps proven secure?
No. The verification covers a Linux test build, not the mobile apps.
Are group chats covered?
No, the verification covers only two-party conversations.
Sources
- À la recherche de la preuve parfaite : Google a lâché 80 agents IA sur le code de Signal pour prouver qu’il était sûr, Numerama,
- https://deepmind.google/blog/introducing-alphaproof-code/, deepmind.google, (primary source)