DeepMind Proved Signal’s Encryption

Google DeepMind has mathematically verified the open-source libsignal library, the core of the Signal protocol, using an agent framework called AlphaProof Code. This library, used by Signal, WhatsApp, Meta Messenger, and Google Messages, was proven all the way down to compiled machine code in just four weeks—a process that would normally take experts years.

The team produced 6.5 million lines of machine-checked proof in the Lean proof language. The agents are powered by frontier models like Gemini Carbon. DeepMind also shared the verification tools and technical papers with the community as open source.

Not the source code, but the running machine code was verified

Previous formal verification efforts were usually limited to examining protocol schemes or high-level source code. AlphaProof Code, however, targeted the AArch64 binary machine code running on user devices, factoring compiler bugs into the equation as well. In other words, the proof shows not just that the design is correct on paper, but that the code delivered to the device carries the same guarantees.

What guarantees were proven

Three points stand out in the study:

• It was shown that forward secrecy and post-compromise security guarantees are fully preserved in the compiled code.
• The mathematical consistency of the PQXDH handshake and the Triple Ratchet algorithm was verified for post-quantum resilience.
• Only 1 percent of the codebase was flagged for secondary human review; throughout the process, known historical bugs and artificial vulnerabilities planted for testing purposes were caught.

Support came from the Signal side as well

Signal CTO Ehren Kret welcomed the work, saying that at a time when AI-powered cyber threats are rising rapidly, automated theorem-proving methods reinforce confidence in the protocol’s privacy guarantees.

Researchers expect the same method could become standard in areas where the margin for error must be close to zero, such as medical device software, aviation control systems, and national power grids.

What changes in practice

The discussion has now moved beyond the question of “can AI write code?” The real question is whether AI can mathematically guarantee the correctness of a critical system. Having the proofs checked by the Lean kernel removes differences in human interpretation and lines that might otherwise be overlooked. The timing is striking too: reducing an effort measured in years to four weeks could directly affect how quickly security updates are verified.

Source: WebTuring