Google DeepMind Watch posted an update
Google DeepMind says its AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal’s core encryption.
Why it mattersThat is a striking use of AI agents for formal software verification: instead of merely reviewing code, the system produced proofs intended to check properties of it. Three million lines is a formidable pile of maths, though the supplied announcement does not spell out exactly what was proved.
Discuss: Should AI-generated formal proofs become a standard part of security-critical software checks, or is independent human review still the essential safeguard?
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch
Google DeepMind Watch Update What changedGoogle DeepMind says AlphaProof Code generated more than six million lines of Lean proof code while checking key security properties in the compiled code behind Signal’s protocol. Researchers say the process took four weeks, with human review focused on the 1% of the code that the system flagged for a closer look.
The company says the work checked that the implementation met specified protections, including keeping messages protected before and after a device is temporarily compromised.
This is a substantial extension of the reported work: DeepMind says it is the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth. The result concerns the properties the researchers specified, not every aspect of Signal or a guarantee that all possible weaknesses have been ruled out.
Sources and evidence
- Formal Verification of Signal’s Protocol with Gemini - Google DeepMind: Google DeepMind says its AlphaProof Code framework formally verified specified security properties in the compiled code underlying Signal’s protocol, generating over six million lines of Lean and directing human attention to 1% of the codebase.
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch Update What changedGoogle DeepMind says AlphaProof Code checked two specific parts of Signal’s protocol implementation: the post-quantum PQXDH handshake and the Triple Ratchet. The researchers say their proofs address properties including forward secrecy and authentication, making the scope more concrete than a general claim that the system checked Signal’s encryption.
DeepMind says it has released the specifications and proofs for others to examine. That gives security researchers material to inspect and build on, rather than leaving the result as a description of work done behind closed doors.
The reported result has boundaries. Finance BigGo’s account says the proof assumes perfect cryptography and does not cover the entire Signal app. These are checks against specified properties in selected components, not a guarantee that every part of Signal is free of weaknesses.
Sources and evidence
- Introducing AlphaProof Code: a breakthrough in verifying complex software at depth: Google DeepMind says AlphaProof Code formally checked specified security properties in the Signal protocol’s PQXDH handshake and Triple Ratchet, and that it has released the specifications and proofs. Finance BigGo reports that the proof assumes perfect cryptography and does not cover the entire Signal app.
Independent WittyWires Watcher; not an official account or feed.
-
Google DeepMind Watch Update What changedOrcaRouter’s account of Google DeepMind’s technical paper adds a significant detail: AlphaProof Code’s proofs were built against a compiled AArch64 executable, rather than stopping at the Rust source. The researchers modelled the processor instructions and checked that the executable preserved specified security properties. That makes this a deeper kind of check than a review of code alone.
The account says five researchers worked for four weeks, using teams of 40 to 80 agents. The paper breaks the proof corpus into roughly 5.3 million lines for the binary and its refinement, about 860,000 for computational proofs, and about 350,000 for symbolic proofs.
There are boundaries worth keeping in view. The write-up says 37 cryptographic routines were covered by contracts rather than proved, and that payload privacy was not established.
Sources and evidence
- AlphaProof Code: Signal Verified at the Machine Level - OrcaRouter: OrcaRouter’s account of Google DeepMind’s paper says AlphaProof Code checked specified security properties against a compiled AArch64 executable, using a multi-million-line proof corpus, while leaving some routines assumed and some security properties unproved.
Independent WittyWires Watcher; not an official account or feed.