Anthropic Puts Claude To Work Proving Software Is Safe

Anthropic announced on October 23, 2026, that its Claude model has generated mathematical proofs confirming that existing software code operates correctly. The company presented the development as a step toward trusting AI-generated software, though it acknowledged significant limitations on the scope of these guarantees. Anthropic’s Frontier Red Team partnered with the company Theorem to shift the proof work to Claude. In a matter of weeks, the model generated approximately 180 million lines of mathematical proof code.

The work focused on test programs drawn from 123 open-source libraries within Google’s OSS-Fuzz security project. Historically, formal verification, the process of proving software runs without error under all conditions, has been prohibitively expensive. Specialists required 20 person-years to verify seL4, a small operating system kernel. Because of this cost, the technique has largely been restricted to extreme cases such as Airbus avionics software. Claude did not write new software for this project. Instead, it verified existing code by turning crash tests into mathematical problems, analyzing compiled machine code, and constructing extensive proofs.

The generated proofs are checked automatically by a separate verification program called Rocq. Anthropic stated that this process requires almost no human reviewers. The researchers describe the proof generation as fully autonomous, while noting that the tools and formal foundations were built by humans at Theorem. During this work, Claude encountered hundreds of bugs and was permitted to fix them. The researchers reported that many of these issues likely would have escaped conventional testing because they only appear under narrow conditions, such as unusually large input files or a specific sequence of accesses.

The effort produced specific security results. In an attempt to verify a widely used PDF reader, the team found as many security bugs in one week as are usually reported for that software in a full year. In another application, Claude identified a logical error that had gone unnoticed for more than a decade. Anthropic noted that even when a complete proof fails, the attempt can still reveal real bugs. The company is careful to state that formally verified does not mean error-free. The proofs hold only under defined conditions, including a particular processor architecture, single-threaded programs, and inputs smaller than 64 MB.

Real-world deployments of these libraries often run on different hardware and configurations. Cost remains a constraint for the project. Individual proofs consume hundreds of millions of tokens, and Anthropic has not published total project costs. A smaller test run indicated that proofs can be reused. Efficiency is improving as well: Opus 5.5 solves half the proofs using about 40 percent fewer tokens than Opus 5. Anthropic expects formal verification with AI to become as practical as standard testing. Theorem’s roadmap already targets critical infrastructure, including glibc, OpenSSL, and cURL.