Google DeepMind’s AI agents formally verify Signal’s encryption code

What happened

Google DeepMind used its AI agents to formally verify the core security properties of the Signal encryption protocol using a new framework called AlphaProof Code. The verification process focused on the open-source libsignal library, which underpins encryption in apps like Signal, WhatsApp, Messenger, and Google Messages. Over four weeks, teams of AI agents generated 6.5 million lines of mathematical proof, confirming features such as forward secrecy and authentication down to the machine code level. The results and specifications are being released publicly for further review.

Key details

• AlphaProof Code is the new framework developed by Google DeepMind for this verification.
• The AI agents produced 6.5 million lines of proof in four weeks.
• Verification covered one-to-one chats in a custom build of the libsignal library, not the full Signal apps.
• The process involved 40 to 80 AI agents running on Gemini 4 Carbon, with specifications reviewed by humans.
• The verification did not include group messaging or key generation, and assumes cryptographic routines work as specified.

Why it matters

Developers and users of apps relying on the Signal protocol, such as Signal, WhatsApp, Messenger, and Google Messages, gain increased assurance that the core encryption code meets its security claims. Organizations can now access the published proofs and specifications to independently verify the results or build on them. The verification does not cover all aspects of the protocol or the full apps, so some risks remain unaddressed, but this step reduces uncertainty about the security of one-to-one encrypted messaging.