Google DeepMind announced AlphaProof Code on 10 November 2026, a multi-agent framework for formal software verification. Five researchers, working with Gemini 4 Carbon agent teams over four weeks, produced roughly 6.5 million lines of machine-checked Lean proof code for Signal's unmodified libsignal library (v0.105.4), compiled to a statically linked AArch64 Linux executable.
The proofs cover the post-quantum PQXDH handshake and the Triple Ratchet, and establish four properties at machine-code level: forward secrecy, authentication, post-compromise security and handshake secrecy. Campaigns typically ran 40–80 agents on one machine for 24–72 hours, and the proofs rest on about 26,000 lines of trusted specification.
The team validated its specifications by reverse-applying two real libsignal bug fixes — in both cases the proofs correctly detected the security failure. For scale: formally verifying the seL4 microkernel took 20 person-years; this took five people four weeks with agents.
The limits matter as much as the headline, and the researchers state them themselves. The proofs concern one-to-one sessions only, assume ideal cryptography, and do not certify the full Signal app. And the computational proof gives no guarantee against a quantum attacker.
Key facts
| Item | Details |
|---|---|
| Announcement | AlphaProof Code, 10 November 2026 (Google DeepMind) |
| What it is | Multi-agent framework for formal software verification |
| Target | Signal's unmodified libsignal library, v0.105.4, compiled to a statically linked AArch64 Linux executable |
| Proof output | ~6.5 million lines of machine-checked Lean proof code |
| Team and time | Five researchers with Gemini 4 Carbon agent teams, over four weeks |
| Campaigns | Typically 40–80 agents on one machine for 24–72 hours |
| Trusted specification | ~26,000 lines |
| Properties proved (machine-code level) | Forward secrecy, authentication, post-compromise security, handshake secrecy |
| Coverage | Post-quantum PQXDH handshake and Triple Ratchet |
| Validation | Two real libsignal bug fixes reverse-applied — proofs correctly detected the security failure both times |
| Scale comparison | seL4 microkernel verification took 20 person-years |
| Limits (stated by the researchers) | One-to-one sessions only; ideal cryptography assumed; does not certify the full Signal app; no guarantee against a quantum attacker |
What this means, and what to watch
Formal verification has long been one of the most labour-intensive things in software — seL4's 20 person-years is the classic reference point. AlphaProof Code suggests agent teams can compress that work dramatically for a well-specified target. But the limits are part of the result, not a footnote: a proof that covers one-to-one sessions, assumes ideal cryptography, and says nothing about a quantum attacker is a real achievement with a defined boundary. The next question is whether the same approach extends to wider targets — group sessions, full applications, and systems where the specification itself is the hard part.
A note on verification
The facts above are verified from the sources named in the research file: The Register, tech-insider.org, aiincontext.com and the Alitheia Project. The limits in this post are those stated by the researchers themselves, as reported in the coverage; no figure has been added or adjusted.
FAQs
What is AlphaProof Code?
A multi-agent framework from Google DeepMind for formal software verification — using teams of AI agents, working with human researchers, to produce machine-checked mathematical proofs that software behaves as specified.
What exactly was proved about Signal?
The proofs cover Signal's unmodified libsignal library (v0.105.4): the post-quantum PQXDH handshake and the Triple Ratchet. They establish four properties at machine-code level — forward secrecy, authentication, post-compromise security and handshake secrecy.
Does this mean the whole Signal app is certified secure?
No. The researchers state the limits themselves: the proofs concern one-to-one sessions only, assume ideal cryptography, do not certify the full Signal app, and give no guarantee against a quantum attacker.
How did the team check the proofs were meaningful?
By reverse-applying two real libsignal bug fixes — reintroducing known bugs — and confirming the proofs correctly detected the security failure in both cases.
Source: Google DeepMind's AlphaProof Code announcement (10 November 2026), as reported by The Register, tech-insider.org, aiincontext.com and the Alitheia Project. Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!