Editorial illustration for DeepMind's AlphaProof Code Verified Signal's Cryptographic Core in Four Weeks
DeepMind's AlphaProof Verifies Crypto Code in Weeks
Verifying that software actually does what its specification promises used to be a job measured in years, not weeks. Google DeepMind says it has cut that timeline down to a month, using a new system called AlphaProof Code. The AI tool mathematically proves that code matches its security requirements, line by line, rather than just checking the design on paper.
DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri picked a hard target to prove the point: Signal's encryption protocol, the cryptographic core that protects messages on millions of phones. Past verification attempts in this space tended to stall at the theoretical stage, confirming that a design was sound without ever touching the code running on actual devices. AlphaProof Code went further, working directly on Signal's deployed implementation.
The process started with experts writing out formal security requirements, then handed off to AI agents built on Gemini Carbon, which produced millions of lines in the Lean proof language to check the work. DeepMind is framing this as a first step toward applying the same method to other systems where a hidden bug carries real consequences, from power grids to medical devices.
Earlier verification efforts stopped at the theoretical design stage, according to a blog post by DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri. AlphaProof Code verified the actual cryptographic core code running on users' devices. After experts defined formal security requirements, AI agents built on Gemini Carbon wrote over six million lines in the Lean proof language, flagging roughly one percent of the codebase for human review.
Why this matters
For developers and founders betting on secure messaging, this is the first sign that formal verification might actually scale to real production code rather than staying a research curiosity. Signal's cryptographic core isn't some toy example; it's the thing billions of messages pass through daily. Compressing a years-long proof effort into four weeks changes the economics of security auditing. Teams that could never justify the cost of formal verification might now afford it.
We'd still want to see this replicated outside DeepMind's own blog post before treating it as settled. One case study, even a high-profile one, isn't proof the method generalizes to messier codebases without Signal's relatively clean architecture. Kohli and Chaudhuri's framing, that earlier tools stopped at theoretical design while AlphaProof Code checked the actual device-running code, is the real claim worth testing.
If independent researchers can point AlphaProof Code at other widely-used encryption libraries and get similar results, that's when this stops being a demo and starts being infrastructure. Watch for who tries next.
Related Reading
Browse all Google news →
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives US data brokers sell USD 500M in RL training environments to Chinese AI labs
From Archives Grok Build coding agent leaks confidential data, some belonging to SpaceXSI
From Archives Gemini Decisions API cuts costs nearly 5x by billing zero output tokens
LLMs & Generative AI
From Archives Google launches Gemini 4 Carbon, claiming cybersecurity gains over rivals
Business & Startups
Common Questions Answered
How does AlphaProof Code differ from traditional software verification methods?
AlphaProof Code mathematically proves that actual code matches its security requirements line by line, rather than just checking the design on paper like earlier verification efforts. This approach reduces verification timelines from years to approximately four weeks, making formal verification practical for real production code instead of just theoretical designs.
What specific achievement did DeepMind accomplish by verifying Signal's encryption protocol?
DeepMind used AlphaProof Code to verify Signal's cryptographic core—the actual code running on users' devices that processes billions of messages daily—in just four weeks. AI agents built on Gemini Carbon wrote over six million lines in the Lean proof language and flagged roughly one percent of the codebase for human review, demonstrating that formal verification can scale to real-world applications.
How does the speed improvement of AlphaProof Code impact the economics of security auditing?
By compressing a years-long proof effort into four weeks, AlphaProof Code dramatically reduces the cost and time required for formal verification. This makes security auditing economically feasible for development teams and founders who previously could never justify the expense of formal verification, potentially transforming how cryptographic systems are validated.
What role did AI agents play in the verification of Signal's cryptographic core?
AI agents built on Gemini Carbon automatically wrote over six million lines of code in the Lean proof language to verify Signal's encryption protocol. After human experts defined the formal security requirements, these AI agents performed the bulk of the verification work and identified approximately one percent of the codebase that required human review.
Further Reading
- Papers with Code Benchmarks - Papers with Code
- Chatbot Arena Leaderboard - LMSYS
