Research4 sources35 comments in the community

Introducing AlphaProof Code: a breakthrough in verifying complex software at depth

The Brief · written from 4 sources

Google DeepMind used its AlphaProof Code agents to verify Signal's core encryption, writing three million lines of proofs in Lean.
Wired reports the work was a four-week proof-of-concept, and The Register says it links protocol-level security guarantees to a widely used cryptographic implementation.
DeepMind says human experts worked with the AI agents to decide what to prove.

Why it matters

It suggests AI-assisted formal verification may be able to scale to large, real-world codebases, although the work is described as a proof-of-concept.

Who’s affected

Signal developers
Teams working on formal verification of cryptographic software
Developers interested in AI-assisted code verification

#formal-verification
#signal
#cryptography
#ai-agents
#lean

AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal’s core encryption.

Read the full story at deepmind.google ↗

What developers are saying

Based on 35 comments · Mixed reactions

Hacker News commenters were largely impressed that Signal's code could be formally verified and saw AI-assisted verification as a hopeful sign for software safety. Skeptics noted that the proof appears to check the implementation rather than the specification, asked how Lean proofs are tied to the actual code, and questioned the cost and whether AI could take over the human decision-making.

Impressive verification achievement30%
Hope for safer AI-written software20%
Spec versus implementation gap20%
Proof-to-code correspondence doubts15%
Cost and limits of AI15%

In favour

+ Shows that a large, real codebase like Signal's can be checked under formal methods
+ Could make software safer as AI takes on more development work

Concerns

− The proof seems to cover the implementation, not the specification
− It is unclear how a Lean proof shows that it represents the code it proves
− The cost of the effort is unknown
− Human experts still decide what to prove, which commenters see as a limit of AI