NEW DELHI: Google DeepMind has achieved a major milestone in cybersecurity by using artificial intelligence to formally verify key security properties of the Signal messaging protocol, which also underpins the encryption used by WhatsApp.

The work was carried out using a new AI framework called AlphaProof Code, which helped researchers examine the cryptographic implementation down to the machine-code level. The project was completed in around four weeks during October 2026, with AI agents generating more than six million lines of mathematical proof.

The development could strengthen confidence in the technology protecting private digital conversations. However, it does not mean that every part of WhatsApp or Signal has been declared completely free of security vulnerabilities.

How Google DeepMind used AI to verify encryption

Google DeepMind research scientist Swarat Chaudhuri shared details of the achievement, explaining how the team used AlphaProof Code to formally verify the Signal cryptographic protocol.

Formal verification is a mathematical method of checking whether software meets specific security and functional requirements. Unlike conventional testing, which looks for errors by running software under different conditions, formal verification attempts to establish mathematically that selected properties hold.

The researchers translated security requirements into formal specifications using Lean, a programming language and proof assistant. AI agents then generated mathematical proofs, which were checked by Lean to establish their validity.

According to the report, the team generated more than six million lines of machine-checked Lean proof over four weeks in October.

The verification extended beyond the protocol’s design to its actual production implementation, including compiled machine code for ARM’s AArch64 architecture. This is significant because machine code contains the instructions that a device’s processor ultimately executes.

What parts of Signal’s security were examined?

The researchers focused on important security properties of Signal’s cryptographic implementation. These include protections that help maintain the confidentiality and authenticity of digital conversations.

Some of the key areas examined were:

  • Forward secrecy: Helps protect previously exchanged messages even if a user’s current encryption keys are compromised.
  • Authentication: Helps ensure that users are communicating with the intended person.
  • Post-compromise security: Helps restore protection after a temporary compromise of encryption keys.
  • PQXDH handshake: A key-establishment mechanism that includes protection designed with future quantum-computing threats in mind.
  • Triple Ratchet: A cryptographic mechanism that manages and updates encryption keys to maintain conversation security.

The work examined important components of the protocol rather than every feature of the messaging applications.

This distinction matters because verifying a cryptographic implementation does not automatically establish that the entire application, its devices, operating systems or user accounts are immune to attacks.

AI agents successfully detected previously fixed bugs

Google DeepMind also tested whether AlphaProof Code could identify security problems in modified software.

Researchers deliberately reintroduced two bugs that had previously been fixed in Signal’s code. The AI-assisted verification system successfully detected that the altered versions no longer satisfied the required security properties.

This test was important because it demonstrated that the framework could identify changes that weakened security rather than merely generate proofs for software that was already functioning correctly.

The project combined AI-generated reasoning with mathematical verification. Instead of relying only on an AI model’s explanation, the researchers used Lean to check whether the generated proofs were mathematically valid.

The approach could help developers identify certain vulnerabilities earlier and reduce the amount of manual effort traditionally required for formal verification.

Does this mean WhatsApp and Signal chats are completely unhackable?

No. The achievement is an important advance in verifying the security of the underlying cryptographic code, but it should not be interpreted as proof that every WhatsApp or Signal conversation is completely immune to hacking.

End-to-end encryption is designed to ensure that only the intended participants can read messages. WhatsApp says personal messages and calls are protected using end-to-end encryption based on the Signal protocol, meaning neither WhatsApp nor Meta can read those personal conversations.

However, encryption is only one part of digital security. A person’s account or device can still be exposed through phishing, malicious software, stolen credentials, compromised backups or physical access to an unlocked phone.

Similarly, the verification conducted by Google DeepMind covered selected security properties of Signal’s cryptographic implementation. It did not amount to a complete security audit of every feature of Signal or WhatsApp.

Users should therefore continue following basic precautions, including keeping their devices and applications updated, enabling available account security features and avoiding suspicious links or attachments.

Why the achievement matters for cybersecurity

Formal verification has traditionally required substantial expertise in mathematics, cryptography and software engineering. Examining complex software at such a detailed level can take years of work, making the process difficult to apply across large technology systems.

Google DeepMind believes AI can make this process faster and more practical by assisting researchers in generating mathematical proofs.

The technology could eventually be applied to other critical software systems, including medical devices, financial infrastructure and power networks. Security failures in these areas can have consequences extending well beyond individual users.

The company has made its formal specifications and verification work available for the open-source community to inspect and build upon.

Human experts remain essential

Despite the progress made by AlphaProof Code, human researchers continue to play an important role in the verification process.

Experts must decide which security properties need to be examined, establish the assumptions behind the verification and ensure that the mathematical specifications accurately reflect the intended security requirements.

AI can help generate proofs and accelerate the process, but researchers must still determine whether the right questions are being asked and whether the results are meaningful.

The Signal project demonstrates how AI-assisted mathematical reasoning could make rigorous software verification more accessible.

For billions of people who use encrypted messaging services, the development is an encouraging step towards stronger digital security. However, it is best understood as a significant verification milestone, not a guarantee that private chats or the applications carrying them can never be compromised.