Key takeaways
- Google DeepMind said AlphaProof Code produced machine-checked Lean proofs for the compiled libsignal code behind one-to-one Signal chats, covering the PQXDH handshake and the Triple Ratchet.
- Humans wrote and reviewed the security specifications while agents powered by Gemini Carbon wrote about 6.5 million lines of proof in roughly four weeks, according to the paper.
- Group messaging, multi-device state, sealed sender, side channels and quantum attackers in the computational proof are outside the scope of this first round.
In this story 5 sections
Signal's code for encrypting one-to-one chats now has a machine-checked security proof, and AI agents wrote the proofs, Google DeepMind said on Nov. 10. The lab, part of Alphabet (NASDAQ: GOOGL), introduced AlphaProof Code, a system that checks compiled software against formal security statements written in the Lean proof language, and said the full Signal effort took four weeks.
DeepMind called the result "the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth." It released a technical paper with the Google DeepMind blog post and said the verification artifacts are open source.
What AlphaProof Code checked in Signal
The target was libsignal, the library behind Signal's encryption, at version 0.105.4 together with the spqr component at version 1.6.0, according to the technical paper. Instead of reasoning about a simplified model, the team compiled the unmodified Rust source with a standard compiler into a program for 64-bit Arm (AArch64) Linux and proved properties about that binary.
The proofs cover two parts of the protocol:
- PQXDH, the handshake two devices use to agree on keys before a conversation starts.
- The Triple Ratchet, which the paper describes as the Double Ratchet combined with SPQR, so keys keep changing as messages flow.
The properties include forward secrecy, authentication, post-compromise security, secrecy of the handshake, and the confidentiality and authenticity of messages. The paper sets out a symbolic model, in which an attacker controls the network, and a computational model with a concrete security bound. The computational proof considers classical attackers only.
How people and agents split the work
Human experts, working with a mix of AI models, wrote the Lean specifications that define what counts as secure, and people reviewed every one, DeepMind said. Agents powered by Gemini Carbon then produced the proofs. Lean's kernel is the final judge: a proof counts only if the kernel accepts it against an approved specification.
Teams of 40 to 80 agents ran for 24 to 72 hours in each campaign, the paper says. The agents also flagged about 1% of the codebase for a person to review a second time.
Key figures from the paper and blog:
- Proof size: about 6.5 million lines of Lean, per the paper (the blog says more than six million).
- Trusted base: roughly 26,000 lines of specifications and Arm instruction semantics that people must trust.
- Machine model: 467 AArch64 instructions and 31 system calls.
- Routines replaced by contracts: 39 cryptographic and library functions.
By our calculation, that works out to about 250 lines of machine-generated proof for every trusted line people had to write and review (6.5 million divided by 26,000).
A test with known bugs
To see whether the system would catch real flaws, the team put two bugs that Signal had already fixed back into an isolated copy of the code. In both cases, the blog said, the system produced a machine-checked proof that the altered version was insecure. One of them, described in the paper, involved a prekey message that decrypted under an archived session and caused the new session state, including a signed prekey secret, to be stored.
Signal's chief technology officer, Ehren Kret, is quoted in the post: "Using AI agents to translate live code into valid Lean theorems is a major breakthrough." DeepMind also thanked Signal's Rolfe Schmidt and Michael Rosenberg for feedback on the method. Signal had not posted about the work on its own blog as of early Nov. 11.
What the proofs leave out
The paper lists several areas outside this round:
- Group messaging, which it names as future work.
- Multi-device state and sealed sender.
- Post-quantum healing in SPQR.
- Side channel attacks.
- Quantum attackers in the computational proof.
So the result applies to one-to-one sessions on the tested build, not to every part of the Signal app. The work examined existing code, and DeepMind did not describe any change to the app itself.
Where AlphaProof Code goes next
DeepMind said the project builds on CodeMender, its earlier AI agent for code security, and it has posted the Lean development, with specifications and proofs, on its code verification site. The company did not give a date for applying AlphaProof Code to other software or say whether it will become a product.
The news follows other recent security work from AI labs, including the OpenAI model that found 81 Firefox flaws in our GPT-6 Cyber report. Google also set Gemini 4 Carbon pricing this week. More coverage is on our tech page.
Sources
Share this story
Spotted an error? Tell us


