AlphaProof Code verifies Signal at the machine-code level, linking mathematical security guarantees to the compiled implementation of a widely deployed cryptographic protocol. Google DeepMind says its AI-agent framework completed the proof project in four weeks.

The result covers important parts of Signal's core cryptography, not the entire Signal app or every way people use it. DeepMind has published the technical paper, Lean proofs and specifications so researchers can inspect the claims and extend the work.

The AlphaProof Code project establishes four concrete results:

• About 6.5 million lines of Lean proof code were generated.
• The proof targets an unmodified AArch64 libsignal binary.
• PQXDH and Triple Ratchet properties were covered.
• Two historical bugs were detected when reintroduced.

AlphaProof Code Verifies Signal From Specification to Machine Code

DeepMind's announcement describes AlphaProof Code as a multi-agent framework for formal specification and software verification. Human researchers decide what properties matter and review the specifications; AI agents then construct proofs that the Lean theorem prover checks mechanically.

The agents were powered by Gemini Carbon and organized into specialized roles. A specification agent helped translate security goals into Lean statements, prover agents generated the proof, and an auditing stage highlighted areas that needed closer human attention.

DeepMind says the agents produced approximately 6.5 million lines of machine-checked Lean proof code. Those proofs rest on roughly 26,000 lines of trusted specifications and AArch64 semantics, according to the technical paper.

The target was libsignal version 0.105.4 compiled for AArch64 from unmodified source with the standard Rust compiler. By checking the compiled binary rather than only the source, the work seeks to close the gap where compiler behavior or low-level implementation details can invalidate a protocol-level argument.

PQXDH and Triple Ratchet Receive Machine-Checked Proofs

The verified scope includes Signal's post-quantum PQXDH handshake and Triple Ratchet. The researchers proved properties covering authentication, forward secrecy, post-compromise security and handshake secrecy under the assumptions defined in their symbolic and computational models.

Forward secrecy aims to protect earlier messages if a long-term key is later exposed. Post-compromise security addresses whether a conversation can recover protection after an attacker temporarily obtains session state. Both guarantees depend on precise assumptions, so the published specification is as important as the generated proof.

The project also tested whether the formalization could distinguish correct code from known-bad code. Researchers restored two previously fixed libsignal bugs in isolated builds; AlphaProof Code then produced machine-checked evidence that the altered binaries violated the specified security theorem.

That test does not mean the system discovered new Signal vulnerabilities. WIRED reported that no new flaw was found in the examined portions and described the exercise as a proof of concept that built on earlier Signal Foundation verification work.

Background Reading

• GDPx Benchmark Launches With Claude Opus 5.5 Leading at 12%
• CrowdStrike AI Kill Switch Launches in Falcon Guardian

Six Million Proof Lines Still Depend on a Trusted Core

Formal verification proves that code satisfies a stated model; it does not prove that every possible security requirement was stated correctly. AlphaProof therefore keeps human review at the specification layer and relies on Lean's small kernel to reject invalid proof steps produced by the agents.

The trusted computing base also includes the approved specifications, the AArch64 model, certain cryptographic contracts and the reproducible-build tooling. A mistake in those trusted components can affect the meaning of the result even when every generated proof line passes Lean's checks.

The paper lists clear boundaries. The computational result does not cover quantum attackers, multi-device state or all large-scale usage patterns. Its symbolic analysis omits computational secrecy, conversations with more than two parties and some post-quantum healing behavior.

The proof also does not certify Signal's user interface, server operations, phone operating systems or every surrounding dependency. Reporting the result as proof that the entire Signal app is unhackable would go beyond the published evidence.

Open Proof Artifacts Set Up Independent Review

DeepMind has released the Lean development and supporting code through an open repository. That allows cryptographers and formal-methods researchers to inspect the specifications, challenge assumptions and test whether the workflow can be reproduced or adapted to later libsignal releases.

The four-week timeline is the most commercially significant claim. Major verification projects have historically required years of specialist labor, so reliable proof agents could make formal methods practical for more firmware, financial software, medical devices and critical infrastructure.

Cost and generality remain unresolved. Google has not released AlphaProof Code as a commercial product or disclosed what the Signal project cost, while each new system may require experts to define different properties, models and trusted components before autonomous proving begins.

The immediate achievement is therefore narrow but substantial: AI agents helped connect selected Signal security guarantees to deployed-style machine code, with every proof checked by Lean and the artifacts opened for scrutiny. Wider impact will depend on independent review, reproduction and successful proofs for less carefully studied software.

Illustrative feature image: computer literacy training photographed by Christian Yakubu, via Wikimedia Commons under CC BY-SA 4.0. Cropped and resized to 16:9; this derivative remains under CC BY-SA 4.0 and does not show AlphaProof Code, Signal or the DeepMind research.