Can Claude Prove Its Own Code Is Safe? Inside the New AI Verification

According to the-decoder.com, Anthropic and partner Theorem tested Claude not just to write code, but to mathematically prove its correctness. Traditional formal verification required extreme investments—like a team of specialists spending 20 person-years on the small seL4 kernel. Claude changed this baseline by generating roughly 180 million lines of mathematical proof code for 123 open-source libraries in just a few weeks.

How automated code verification works

The process moves far beyond standard unit testing. Claude translates crash tests into mathematical problems, analyzes compiled machine code, and outputs massive proof files. An independent verification program called Rocq then checks these proofs automatically, requiring almost no human reviewers to oversee the math.

While running these verification tasks, Claude encountered hundreds of subtle bugs and was authorized to fix them. During a test on a widely used PDF reader, the model uncovered as many security flaws in a single week as typically surface in an entire year of normal operation. Another test exposed a logical error that had remained hidden for over a decade.

Limits and compute costs in practice

Despite the breakthrough, the proofs come with strict practical boundaries. They only hold under single-threaded execution, specific processor architectures, and inputs under 64 MB. Furthermore, the compute overhead is massive, consuming hundreds of millions of tokens per individual proof.

Vale a pena para a sua rotina?

For development teams handling critical infrastructure, this shift promises to cut down tedious audit cycles and catch edge-case bugs early. However, high token consumption and strict hardware constraints mean full formal verification remains suited for core libraries rather than everyday application features.