Google DeepMind Uses AI to Verify Signal's Cryptographic Code

On November 10, 2026, Google DeepMind presented the results of a study in which AI agents helped mathematically verify key security mechanisms of the Signal cryptographic protocol. For this work, scientists developed the AlphaProof Code system, which combines large language models, formal verification tools, and oversight by specialists. According to the researchers, they succeeded for the first time in linking the security proofs of a widely used cryptographic protocol with the machine code of its unmodified implementation.

The project took four weeks and required the participation of five researchers. During this time, agents based on Gemini 4 Carbon created about 6.5 million lines of proofs in the Lean language, which were verified by the mathematical core of this system. The results were published along with technical documentation and source code so that other specialists could study and reproduce the verification.

Which Signal security mechanisms were verified by researchers

The main focus of the study was the libsignal library, which implements the Signal cryptographic protocol. This protocol is used for end-to-end message encryption not only in the eponymous messenger but also in other communication services, including WhatsApp and Google Messages. The researchers examined two important mechanisms: PQXDH, responsible for secure key exchange during connection establishment, and Triple Ratchet, which updates cryptographic keys as the conversation progresses.

The verification covered several security properties. Among them were the confidentiality and authenticity of messages, protection of previous correspondence in the event of subsequent key compromise, and the ability to restore the security of further exchanges after certain types of compromise. The researchers also analyzed specific security properties of the connection establishment procedure, considering potential quantum threats.

Unlike traditional verification of a mathematical model of a protocol, DeepMind studied the compiled implementation directly. For this, the researchers used version 0.105.4 of libsignal and examined the executable code for the AArch64 architecture used in Arm processors. This approach allows for verification not only of the cryptographic scheme but also of how well its software implementation adheres to the specified security properties.

How AlphaProof Code uses AI to verify programs

Formal verification differs from regular testing in that instead of checking individual scenarios, developers mathematically prove the fulfillment of certain requirements. For example, a condition can be formulated where an unauthorized participant should not gain access to the content of a message, and then it is checked whether this condition holds in all cases provided by the model.

In AlphaProof Code, this process begins with creating a formal specification—a precise mathematical description of the properties to be verified. AI agents help translate security requirements into expressions in the Lean language, but the final formulations are checked and approved by humans. After this, another group of agents generates proofs, which the Lean system verifies independently of how they were obtained.

During the study, about 6.5 million lines of machine-verified proofs were prepared. They relied on approximately 26,000 lines of trusted specifications and descriptions of the processor architecture's operation. The system also highlighted areas requiring additional human verification, allowing researchers to focus on the most critical assumptions.

To test AlphaProof Code's ability to detect real issues, the team reintroduced two previously fixed vulnerabilities into separate versions of the code. In both cases, the system was able to mathematically demonstrate the violation of the specified security properties. This experiment showed that the tool is capable not only of confirming the correctness of an implementation but also of identifying certain errors.

Does this mean Signal is fully secure

Despite the results, the study does not cover the entire Signal system. The researchers verified a specific implementation of the cryptographic protocol for message exchange between two participants, not all functions of the mobile application. Group chats, specific features of device and server operations, as well as some types of software attacks, were outside the scope of this work.

Moreover, the mathematical proofs are valid under certain assumptions about the security of the cryptographic algorithms used. The researchers applied two models: a symbolic one, where cryptographic operations are considered in an idealized form, and a computational one, taking into account the capabilities of attackers under given conditions. However, the computational proof does not extend to quantum adversaries, and some post-quantum properties of Triple Ratchet still require separate verification.

The researchers did not find new vulnerabilities in the verified components of the current implementation. However, this does not rule out problems in other parts of the application and does not guarantee system security under all circumstances.

Why the study is important for software development

According to Google DeepMind, AlphaProof Code could significantly reduce the labor required for the formal verification of complex software. Previously, such projects took years of work by specialists: for example, the formal verification of the seL4 operating system microkernel required about 20 person-years, and the development of the verified CompCert C compiler took six years.

The most labor-intensive part of the new approach was not the generation of proofs but determining which properties of the software needed to be verified. Human-AI collaboration accelerated the preparation of proofs but did not eliminate the need for expert evaluation of mathematical models and their limitations.

In the future, Google DeepMind plans to apply similar methods to other complex software systems, including banking software and critical infrastructure. However, the possibility of scaling the technology to larger codebases, the cost of computations, and the amount of necessary human oversight remain open questions.