AI Proves Software Safety Without Human Review, Closing a Long-Standing Gap

The history of software has been a search for safety, and now Anthropic says its Claude model has taken a significant step toward writing code while simultaneously proving its correctness. In a development announced on October 23, 2026, Anthropic has deployed Claude to mathematically verify existing software code, a move designed to address the reliability of artificial intelligence-generated programs.

The numbers behind the effort are striking. Anthropic's Frontier Red Team partnered with Theorem to let Claude handle proof work, and the model produced roughly 180 million lines of proof code. This was accomplished in a matter of weeks, drawing on test programs from 123 open-source libraries associated with Google's OSS-Fuzz security project. The effort to mathematically prove that software runs error-free under all circumstances has historically been prohibitively expensive. In one instance, a team of specialists required 20 person-years to formally verify the small operating system kernel seL4, limiting the method to extreme cases such as avionics software for Airbus.

Claude did not write new code during this process. Instead, it verified existing software by translating crash tests into mathematical problems and analyzing compiled machine code. The model generated extensive proofs which an independent verification program, called Rocq, automatically checks. Researchers describe this as fully autonomous proof generation, noting that the tools and formal foundations were constructed by humans at Theorem. While generating these proofs, Claude encountered hundreds of bugs and was authorized to repair them itself. Many of these flaws would have escaped conventional testing because they only surface under specific conditions, such as unusually large input files or a particular sequence of accesses.

The researchers state that during an effort to verify a widely used PDF reader, the team identified as many security bugs in one week as are typically reported for that software in an entire year. In another application, Claude detected a logical error that had remained undetected for more than a decade. Anthropic says that even when a complete proof fails, the attempt still surfaces real bugs. However, the company acknowledges that formally verified does not mean error-free. The proofs only hold under strict conditions involving a specific processor architecture, single-threaded programs, and inputs under 64 MB. In practice, these libraries run on different hardware and in different configurations.

The compute cost for these operations is steep, with individual proofs running up hundreds of millions of tokens. Anthropic does not report total project costs for this initiative. However, if AI can verify its own code, developers can trust AI-generated software more and spend less time reviewing it. That saves costs not just on verification but on the entire coding process. A smaller test run also indicated that proofs are reusable. Newer models are becoming more efficient, with Opus 5.5 solving half the proofs using about 40 percent fewer tokens than Opus 5.

This project strengthens the case for software automation. If AI can write code and guarantee its correctness under defined conditions, the question shifts from whether AI can program to whether AI can prove that the code it writes is safe. Anthropic expects formal verification with AI to soon become as practical as standard testing, and Theorem's roadmap already targets critical infrastructure like glibc, OpenSSL, and cURL.