Community Trust ScoreVerified
Google DeepMind researchers used AI agents to generate mathematical proofs for Signal’s encryption library. Released Nov. 10, the work comprises roughly 6.5 million lines of machine-checked Lean proof code, produced by five people over four weeks.
More coverage is filed under Technology.
Scope of Verified Protocols
The proofs target an AArch64 Linux executable built with libsignal-protocol v0.105.4 and SPQR v1.6.0, operating in a small command-line harness. Theorems cover one-to-one messaging—specifically session initialization, sending, and receiving—linking the compiled build to the security of the PQXDH and Triple Ratchet protocols.
Verified properties include authentication, forward secrecy—ensuring old messages stay secret if keys are stolen—and post-compromise security. Symbolic analysis relies on idealized cryptographic operations and collision-free executions, while computational analysis bounds classical attacker success under stated assumptions.
Limitations and Exclusions
The verified scope excludes key generation, group messaging, and multi-device state. Computational bounds lose utility at sufficiently large workloads, and the proofs do not cover quantum attackers. Results depend on trust in the AArch64 Linux model, the specifications, the contracts of selected routines, and Lean’s checking kernel.
The symbolic proof establishes handshake secrecy during later Diffie-Hellman breaks. The computational proof bounds classical attackers but ignores quantum threats. The report notes that symbolic analysis uses idealized operations that differ from real-world implementations in certain edge cases.
Recent Updates and Verification
A Nov. 11 GitHub update verified ten more library routines, reducing trusted routine hooks from 37 to 27. While retaining libsignal v0.105.4, the update changes the pinned executable to verify machine code for these ten routines rather than relying on modeled contracts.
Public instructions allow for rebuilding the pinned binary and checking proofs. A Linux-only comparator ensures theorems and definitions match trusted statements by replaying proofs in a fresh Lean kernel. To test the specification, authors reintroduced two fixed libsignal bugs; the resulting binaries violated the symbolic security theorem, proving the specification rejects known regressions.
