Google DeepMind's AI agents produce machine-checked security proofs for Signal's encryption library
Google DeepMind researchers say their AI agents produced mathematical proofs connecting specific security guarantees to a compiled build of Signal's encryption library. The evidence has been published online so others can check it.
The work is described in a technical report dated November 10, 2026, under the name AlphaProof Code. The researchers say the agents generated about 6.5 million lines of machine-checked proof code in the Lean language. A machine-checked proof is a mathematical argument that a computer program has verified step by step, rather than one reviewed only by people.
Key points from the report
- DeepMind reports about 6.5 million lines of machine-checked Lean proof code for a compiled Signal library build.
- Five people worked on the proofs over four weeks, directing agents, choosing reference models and reviewing specifications. The paper lists nine authors.
- The verification target is an AArch64 Linux executable built from libsignal-protocol v0.105.4 and SPQR v1.6.0 inside a small command-line harness.
- The theorems cover session initialization, sending a message, and receiving one. Key generation is not covered.
- A November 11 update on GitHub says ten more library routines were verified, cutting trusted routine hooks from 37 to 27.
What the AlphaProof Code report says
According to the report, AlphaProof Code is a human-guided, multi-agent framework. Humans chose what the agents had to prove. The proofs link the compiled executable to stated properties of PQXDH and Triple Ratchet one-to-one messaging, including forward secrecy, authentication, and post-compromise security.
In plain terms, these properties ask three questions: do earlier messages stay secret if keys are later stolen, are senders really who they claim to be, and can a conversation regain secrecy after an attacker gets in.
The report describes two proof models that give different kinds of assurance. The symbolic analysis relies on idealized cryptographic operations and collision-free executions. The computational analysis limits how successful a classical attacker could be under stated cryptographic assumptions.
What the proofs do not cover
The guarantees apply to the specific libsignal command-line build and its modeled execution environment, not to Signal's complete phone applications. The results also depend on trusted elements: the specifications reviewed by humans, Lean's checking kernel, the AArch64 Linux model, contracts for selected routines, and assumptions about the underlying cryptographic primitives.
Several limits are stated in the report. The symbolic proof establishes handshake secrecy under specified later breaks of Diffie-Hellman, but the computational proof does not cover quantum attackers. Group messaging is outside the scope. The computational model excludes multi-device state, and its bounds stop giving a useful guarantee at sufficiently large workloads.
How others can check the work
The project's public repository provides instructions to rebuild a pinned binary, check the Lean proofs, and run a Linux-only comparator. That comparator checks the theorem statements and their dependent definitions against trusted statements and replays the proofs in a fresh kernel. CryptoSlate reports it has not independently rebuilt and checked the full proof package.
Why this matters
Signal is a widely used private messaging protocol, and libsignal is its implementation library. If the reported results hold up under outside review, they would show AI agents completing a formal security audit of real, compiled code that humans directed but did not write line by line. The report itself also shows the limits of such work: the proofs cover a narrow build under explicit assumptions, and key generation, group messaging, and full phone apps remain outside the verified scope.