Formal Verification: AI Verifies Signal's Crypto Core in Four Weeks

What happened

Google DeepMind's AlphaProof Code has mathematically verified the cryptographic core code behind Signal in four weeks, a task that usually takes years, according to a report from The Decoder.

The verification was carried out by AI agents powered by Gemini Carbon. Those agents wrote more than six million lines of proof code. Only about one percent of that output was flagged for human review, which allowed the team to concentrate expert attention on a small fraction of the work.

DeepMind plans to apply the same method to critical infrastructure, including power grids and medical devices. The Decoder's reporting does not set out a timeline or specific deployment partners for those extensions.

Why it matters

Formal verification has long been one of the most reliable ways to show that software behaves as intended, but it has been slow and expensive, requiring specialist mathematicians and engineers. If AI agents can generate the proofs at this scale and the human effort shifts to reviewing the flagged exceptions, the economics of assurance change. Work that was once rationed to the most critical systems could become practical for a much wider set of products and services.

For leaders in AI and digital transformation, the significance lies in the operating model as much as the technology. Verification moves from a scarce, expert-only step to a managed pipeline in which machines produce the bulk of the output and people govern the risk. Organisations modernising their own systems, and public bodies responsible for essential services, will need to decide where that human review should sit and how much autonomy they are prepared to grant.

By the numbers

• Four weeks to verify Signal's cryptographic core code, as reported by The Decoder.
• More than six million lines of proof code written by AI agents powered by Gemini Carbon.
• About one percent of the generated output flagged for human review.

The Renascence take

The headline will draw attention to speed, but the more significant shift is in where human judgement is applied. Assurance is becoming a triage problem, and the people who matter most are those who decide what gets escalated.

Most readers will focus on the four-week timeline and miss the trust architecture that makes it credible. The principle underneath is that confidence comes from designed checkpoints, not from assuming either the machine or the auditor is infallible. A customer-obsessed operator should map its high-stakes processes now, define the exact conditions that trigger human review, and test those thresholds before automating anything.