Security
Google DeepMind says AlphaProof Code formally verified selected Signal protocol guarantees in the libsignal code in four weeks, with millions of lines of Lean proof and no new flaws found in the parts covered; the scope is limited and the line count varies by source
DeepMind says AlphaProof Code formally verified selected Signal guarantees in libsignal in four weeks. It covers core crypto code, not the whole app.
On November 10, 2026, Google DeepMind announced AlphaProof Code, a multi-agent framework that it says formally verified selected security guarantees of the code behind the Signal protocol, down to machine instructions, in about four weeks. The proofs are written in Lean and checked by machine. The work covers the core cryptographic code in libsignal, not the Signal app, and several figures differ between outlets.
The essentials
- What was verified: DeepMind says its agents proved key properties of the actual libsignal code, including protections before and after a temporary device compromise. ShiftDelete notes the low-level crypto building blocks were treated as assumed-secure contracts rather than proven from scratch.
- Scale and method: DeepMind’s post says the agents wrote over three million lines of Lean proof; StartupHub reports more than six million, and ShiftDelete 6.5 million. Humans reviewed every specification before proofs ran, per DeepMind, and reviewed flagged proof code. We could not reconcile the line counts.
- Results: Dealroom, relaying WIRED, says the team found no new vulnerabilities in the covered portions. On older Signal versions, the system flagged known, long-fixed flaws and showed they could not be verified; DeepMind also reintroduced two fixed bugs into an isolated version and the tool caught them.
- Limits: The work did not cover the whole mobile app. DeepMind ran it free of charge, would not share the cost or say if a commercial product will follow, and says it will release the specification and proof. The “first time” claim is DeepMind’s own.
What a formal proof does and does not give
A machine-checked proof shows that code meets a written specification. If the specification is wrong or incomplete, the proof says nothing about the missing case, which is why humans reviewed every specification. StartupHub puts it as: the prover is fast and the specification becomes the security decision.
Why it matters, and caution
Formal verification has usually been too costly for production cryptography; doing it in four weeks would change that, if it scales at a workable cost, which remains open. The Signal protocol also underpins Google Messages and, in modified form, WhatsApp. We read the announcement extract and relays, not the proofs. Not a security audit.
What is not established
- Whether the proofs cover the whole Signal app: no; the scope is core libsignal code.
- The exact number of lines of Lean proof: three million per DeepMind’s post, six million or more per two outlets; not reconciled.
- The cost and scalability of the approach: DeepMind did not share the cost.
- Independent review of the specifications and assumptions: none found yet; artifacts are to be released.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Announced | November 10, 2026, by Google DeepMind | Google DeepMind |
| Time taken | About four weeks | Dealroom (WIRED relay) |
| Proof size | Over three million lines of Lean per DeepMind; over six million per StartupHub | StartupHub |
| Findings | No new vulnerabilities in the covered portions | Dealroom (WIRED relay) |
| Scope | Core cryptographic code, not the full app; low-level primitives assumed | ShiftDelete |
FAQ
Is Signal now proven secure?
No. Selected guarantees of the core protocol code were verified against specifications, with assumptions; the whole app was not covered.
How many lines of proof?
DeepMind’s post says over three million lines of Lean; two outlets report six million or more. We could not reconcile them.
Will the proofs be public?
DeepMind says it will release the specification and verification artifacts for others to use.
Method and limits
Sources checked on November 10, 2026: Google DeepMind’s post (extract), Dealroom, StartupHub, ShiftDelete, Startup Fortune and RuntimeWire, read as search extracts; the Lean proofs and WIRED’s article were not read directly. Not a security audit.
Sources
- Google DeepMind: Introducing AlphaProof Code: a breakthrough in verifying complex software at depth (verified 2026-11-08)
- Dealroom (WIRED relay): DeepMind's AI formally verifies key Signal protocol guarantees in four weeks (verified 2026-11-08)
- StartupHub: DeepMind proved Signal code with 6M lines (verified 2026-11-08)
- ShiftDelete: AlphaProof Code Verifies Signal Encryption at Machine Level (verified 2026-11-08)
- Startup Fortune: Google DeepMind's AI proves key Signal protocol guarantees in four weeks (verified 2026-11-08)
- RuntimeWire: DeepMind's AlphaProof Code proved selected Signal guarantees in four weeks (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy