Google DeepMind said Tuesday that researchers using a new agentic framework, AlphaProof Code, formally verified key security properties of libsignal, the Signal protocol’s reference implementation, down to compiled machine code. In a blog post and a same-day technical paper, the lab said the proofs cover Signal’s post-quantum PQXDH handshake and Triple Ratchet. Wired and Axios also reported the work.
DeepMind said human experts reviewed formal specifications of Signal’s security goals, then agents powered by Gemini Carbon generated the proofs in Lean, a language used to check that code matches a specification. Axios reported that those agents ran on Gemini 4 Carbon, and the technical paper said the proofs were produced by agent teams powered by Gemini 4 Carbon. The paper says the agents that wrote the specifications ran on Gemini 4 Carbon and Anthropic’s Claude Opus 5.5. Google made Gemini 4 Carbon available to paying users and AI Ultra subscribers this week. DeepMind’s post said the agents generated over six million lines of Lean; the paper put the machine-checked proof code at about 6.5 million lines.
The verification took four weeks, DeepMind said. The paper said the proofs establish forward secrecy, authentication, post-compromise security and handshake secrecy for an AArch64 executable built from unmodified libsignal, wrapped in a small command-line harness. Wired noted that the analysis did not formally verify the entire Signal mobile app and did not find new vulnerabilities in the portions it examined. The paper said the results concern that harness build, not Signal’s own applications.
To test the framework, DeepMind said it “reintroduced two real, previously fixed bugs into an isolated version of Signal’s code.” “In both cases, the system produced a machine-checked proof that the buggy version was insecure,” the lab said. Axios described the same check, saying the framework generated proofs that the altered software was insecure.
Wired reported that DeepMind conducted the Signal analysis free of charge as a research project and that Google declined to share the cost. Axios said the work remains a research project, not a commercially available cybersecurity product. Google also declined to comment on whether it plans a commercial formal verification product or service, Wired reported.
Ehren Kret, Signal’s chief technology officer, said in a statement released by DeepMind that “using AI agents to translate live code into valid Lean theorems is a major breakthrough” and that verifying the protocol implementation “gives us even greater confidence in our code’s privacy guarantees.”