Google DeepMind's AI agents have generated 6.5 million lines of machine-checked proofs to verify security properties in Signal's code. The proofs, covering a specific build of Signal's encryption library, were made public for outside verification. The work focused on session initialization, sending, and receiving messages, but key generation and group messaging were excluded.
The proofs were produced over four weeks by five researchers, with humans directing the work and reviewing specifications. The results connect the compiled executable to security guarantees for PQXDH and Triple Ratchet one-to-one messaging, including forward secrecy and authentication. However, the proofs do not cover quantum attackers or multi-device state.
The project's update on Nov. 11 added verification for ten more library routines, reducing trusted routine hooks from 37 to 27. The public package allows other researchers to rebuild the pinned binary and check the proofs. The authors also reintroduced two previously fixed libsignal bugs to test if the specification rejects known regressions.
While the results provide a concrete way to check the connection between security requirements and executable code, they still require trust in Lean’s checking kernel, the specifications, and the AArch64 Linux model. The project repository offers public instructions for rebuilding and checking the proofs, but CryptoSlate has not independently verified the full proof package.