Google DeepMind has introduced AlphaProof Code, a multi-agent AI framework for formal software verification, and used it to mathematically prove key security properties of the actual code behind the Signal protocol. The company said the work produced about 6.5 million lines of machine-checked Lean proof and took four weeks, a task it said historically takes expert teams years.
The target was libsignal v0.105.4 with spqr v1.6.0, Signal's reference implementation, compiled with a stock Rust toolchain into an AArch64 Linux binary. The proofs cover one-to-one sessions, including the PQXDH handshake, the Triple Ratchet, HKDF, HMAC-SHA256, AES-CBC with HMAC, XEdDSA, X25519 and protobuf serialization. Group messaging is listed as future work.
AlphaProof Code splits the job between people and agents. Human experts choose what to prove and review and approve every formal specification, while autonomous prover agents write the proofs. A gatekeeper accepts a proof only if Lean's kernel checks it against the approved specification. Google DeepMind said the prover agents were powered by Gemini 4 Carbon, with a typical agent team running 40 to 80 agents on a single machine for 24 to 72 hours per campaign.
Key facts
- 01Google DeepMind introduced AlphaProof Code on November 10, 2026, a multi-agent framework for formal software verification.
- 02The team proved security properties of libsignal v0.105.4 compiled for AArch64, covering one-to-one Signal sessions.
- 03The proofs total about 6.5 million lines of machine-checked Lean, resting on about 26,000 lines of trusted specifications and semantics.
- 04DeepMind says the verification took four weeks; prover agents were powered by Gemini 4 Carbon.
- 05Reintroducing two previously fixed libsignal bugs produced machine-checked proofs that the buggy versions were insecure.
The 6.5 million lines of proof rest on about 26,000 lines of trusted specifications and AArch64 semantics, and DeepMind said human review covered about one percent of the codebase. The machine-code model alone describes 467 instructions and 31 system calls in 7,578 lines of Lean.
Google DeepMind called the result the first time the implementation of a widely deployed protocol like this has been formally verified by any method, with or without AI. The accompanying paper frames it against earlier landmark efforts: the seL4 microkernel took twenty person-years to verify, and the CompCert compiler took six.
The proven properties include forward secrecy, post-compromise security, authentication, handshake secrecy and resistance to key-compromise impersonation. To test the system, the researchers reintroduced two real, previously fixed libsignal bugs, including a small-subgroup bug, into an isolated copy of the code. In both cases AlphaProof Code produced a machine-checked proof that the buggy version was insecure.
Signal engineers Rolfe Schmidt, Michael Rosenberg and Ehren Kret gave feedback on the methodology, according to the paper. Google DeepMind said the protocol protects billions of private conversations in apps such as Signal and Google Messages, and the paper notes variants also run in WhatsApp and Messenger.
“Using AI agents to translate live code into valid Lean theorems is a major breakthrough.”— Ehren Kret, Chief Technology Officer of Signal
The paper is explicit about scope. The symbolic guarantees assume perfect cryptography, the computational proofs cover classical adversaries only, and side channels are out of scope. The ratchet theorems show the Triple Ratchet is no weaker than the Double Ratchet, but do not prove the post-quantum healing that SPQR adds.
Google DeepMind is open-sourcing the Lean specifications and proofs on GitHub. The company said verifying Signal's core cryptography lays groundwork for securing other systems, from medical device firmware to power grid software, and that scaling the techniques requires continued research.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
