Google DeepMind researchers say AI agents helped produce machine-checked proofs of security properties in compiled code from Signal’s libsignal protocol implementation. The work is notable not just for the millions of lines of proof generated, but for the careful limits around what those proofs establish.
Google DeepMind Watch analysis
What happened
The researchers used AlphaProof Code to examine the PQXDH handshake and Triple Ratchet components in a compiled libsignal command-line programme for AArch64 Linux. Their paper, published on 10 November, describes a four-week project using Gemini 4 Carbon to generate proofs, with Lean checking that those proofs establish specifications approved by human experts. EdTech Innovation Hub’s report gives the project details.
The researchers also reintroduced two previously fixed bugs into experimental builds. The agents produced proofs showing that those altered builds violated the specified security properties. This was a test of the method using historical bugs, not a report of newly discovered vulnerabilities in current Signal software.
Why it matters
Formal verification can connect claims about a cryptographic protocol to the code that actually runs, rather than stopping at the design on paper. That matters because implementation differences can undermine an otherwise sound design. The work offers a substantial example of AI-assisted proof generation being checked by a separate system, rather than asking readers to take an AI agent’s word for it.
The boundary is important. The proofs cover the researchers’ compiled command-line programme, not the complete Signal applications. Some supporting routines were represented by contracts, meaning their behaviour was assumed rather than independently proved. The specifications also define what is being proved and which assumptions and attacker models apply. The paper says most of the effort went into developing those specifications.
Our read
This is a serious demonstration of AI helping with formal software verification, not a certificate that Signal has been proved invulnerable. The reassuring part is that Lean checks the generated proofs against human-approved specifications. The question still worth asking is whether those specifications capture the right risks and how much of the real-world software remains outside the proof.
What to watch
- Whether the released specifications and proofs attract independent scrutiny or useful contributions.
- Whether the approach can be extended beyond the command-line target to more of the software people actually use.
- Whether future work can scale the method to other security-critical systems without weakening the claims being proved.
Discussion spark: For security-critical software, would you trust AI-assisted proofs more if the specifications were independently reviewed, or is checking the generated proof with Lean the more important safeguard?
Sources and evidence
- Google DeepMind researchers verify Signal code with AlphaProof Code | ETIH EdTech News – EdTech Innovation Hub (16 November 2026, 00:47 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind