Research6 sources138 comments in the community

Google Widens Gemini 4 Carbon Access After AI Agents Verify Signal Code

SecurityWeek · By Eduard Kovacs · 1h ago

The Brief · written from 6 sources

Google DeepMind used its AlphaProof Code agent framework to formally verify core encryption in Signal's libsignal.
The AI agents wrote about three million lines of proofs in Lean, as part of a four-week proof-of-concept.
The work links protocol-level security guarantees to a widely used cryptographic implementation, which the researchers say is a first for AI-assisted proofs.

Why it matters

It shows AI agents can help produce machine-checked proofs for real production cryptographic code, which could make formal verification practical for more software, though human experts still decide what to prove.

Who’s affected

Teams relying on the Signal Protocol or libsignal
Developers interested in formal verification of production code

Gemini 4 Carbon
#formal-verification
#signal
#ai-agents
#cryptography
#lean
#google-deepmind

Google opened Carbon to paying developers and businesses as DeepMind used agents built on it to prove libsignal security guarantees. The post Google Widens Gemini 4 Carbon Access After AI Agents Verify Signal Code appeared first on SecurityWeek .

Read the full story at securityweek.com ↗

What developers are saying

Based on 138 comments · Mixed reactions

Hacker News commenters were largely impressed that AI agents formally verified Signal's cryptography and saw it as a sign of safer software to come. Others questioned whether a Lean proof truly matches the implementation, how the proof model was checked, and whether the work replaces human expertise. Several noted the value of the reintroduced bugs test and the limits of the approach beyond cryptography.

Enthusiasm for AI verification35%
Proof matches implementation?20%
Spec versus implementation gap15%
Role of human experts15%
Scope beyond cryptography15%

In favour

+ The agents verified the libsignal implementation, not just a human-written spec.
+ Researchers reintroduced known bugs to confirm the process catches real issues.

Concerns

− AI-written Lean can cheat, so it needs adversarial review.
− Commenters doubt that a Lean model reliably represents the real code.
− Humans still decide what to prove, which raises questions about the limits of AI.

“It’s nothing short of amazing that a huge codebase like Signal’s is apparently not only secure in a practical sense, but secure even in a strict formal sense.”
p-e-w on Hacker News