DeepMind AI mathematically proves Signal protocol security in four weeks
Google DeepMind used AI agents to formally verify the encryption protocol behind Signal and Google Messages in just four weeks.
Research · Source: Google DeepMind
What happened
Google DeepMind just hit a massive milestone in software security. Their new agentic framework, AlphaProof Code, mathematically proved the security of the Signal protocol. This is the exact cryptography that protects billions of private conversations on apps like Signal and Google Messages. The system verified Signal's post-quantum PQXDH handshake and the Triple Ratchet algorithm. This validates the design of forward secrecy and post-compromise security within the libsignal implementation.
The process blended human expertise with artificial intelligence. Human researchers translated security goals into formal mathematical specifications. Then, AI agents powered by the Gemini Carbon model took over the heavy lifting. These agents generated 6.5 million lines of Lean code. Lean is a computer language used to mathematically prove the accuracy of software.
The agents then audited their own proofs. They flagged just one percent of the codebase that required a second look by a human. The entire verification took a five-person team just four weeks. Historically, verifying systems like this takes expert teams years of manual labor. To prove the system works, researchers fed it an isolated version of Signal code containing two previously fixed bugs. The AI successfully produced a machine-checked proof that the buggy version was insecure.
Key facts
- 4 weeks — Time it took a five-person team to verify the protocol
- 6.5 million — Lines of Lean code generated by Gemini Carbon agents
- 1 percent — Portion of the codebase flagged by AI for human review
- 20 person-years — Time it historically took to formally verify the seL4 microkernel
Why it matters
Formal verification is the absolute gold standard of software security. It means mathematically proving that code runs correctly without flaws. But it is usually too expensive and slow for practical use. For context, formally verifying the seL4 microkernel took twenty person-years. Developing the verified CompCert C compiler took six years. DeepMind just collapsed that timeline to weeks. This means elite security guarantees are no longer restricted to massive tech monopolies or decade-long academic projects.
Real-world vulnerabilities rarely exist in the high-level design of a cryptographic protocol. They hide in the messy gap between that design and the actual compiled code running on a device. Most protocol verification stops at the design level. DeepMind pushed this down to the machine-code level. As AI accelerates offensive cyber threats, we need automated defense to keep pace. The bottleneck is now shifting from writing complex mathematical proofs to simply defining the security specifications.
For builders
Automated formal verification is now accessible
DeepMind is open-sourcing their formal verification artifacts for community use. Security engineers can study these to build automated theorem-proving pipelines for their own critical systems. Companies that build medical device firmware or power grid software will soon be expected to provide mathematical proof of their security.
Specification is the new engineering bottleneck
AI agents can now write the millions of lines of Lean code required for a proof. But human experts still need to define what exactly needs to be proven. Founders building developer tools should focus on human-AI collaboration interfaces that make software specification easier.
Defensive AI must match offensive scale
Offensive AI capabilities are rising rapidly. Insecure systems worldwide will be exploited at an unprecedented scale. Builders must integrate automated defense agents like DeepMind's CodeMender into their pipelines to counter these automated threats.
My take
Formal verification used to be a luxury for academics with decades to spare. Now it is an engineering standard you can execute in a month. If you build critical infrastructure and do not use AI to mathematically prove your code, you are choosing negligence. The math does not lie.
Original reporting: Google DeepMind. This is my rewrite and opinion.