What was already known
Earlier analyses proved security properties of abstract Signal protocol models, including a formal analysis of PQXDH. The new proofs connect the modeled properties to a compiled reference implementation.
AI involvement
The authors credit Gemini Carbon agents with synthesizing the Lean proofs. Researchers guided the specification work, approved the statements and reviewed selected proof areas.
Importance
The proofs link a deployed protocol’s security model to concrete machine-code behavior, including how the implementation stores session state and handles messages. The authors report that Lean also proved violations for two historical bugs reintroduced into the build.
Technical verification
The paper reports that Lean checked the proofs and that a comparator matched them to held-out specifications. The checked conclusions depend on the stated cryptographic assumptions, routine contracts and AArch64 semantics.
Formal checking
The paper reports that Lean checked the proofs and that a comparator matched them to held-out specifications. The checked conclusions depend on the stated cryptographic assumptions, routine contracts and AArch64 semantics.
- Checker
- Lean kernel and Lean comparator as reported by Chaudhuri and colleagues
- Scope
- AlphaProof Code agents proved security properties of the PQXDH handshake and Triple Ratchet messaging for a command-line harness built from unmodified libsignal 0.105.4 and SPQR 1.6.0 on AArch64 Linux. The symbolic and computational proofs cover different properties and attacker models.
- Basis
- Reported by the linked checker
- Google DeepMind’s Lean development for the Signal buildChecking scope
Human understanding
The authors explain how the proof connects binary traces to symbolic events and give the computational proof structure. The report does not document a human review of every generated argument.
Correction and publication history
New curated research with documented central AI work.
Read this published revision