AI Model Generates 180 Million Lines of Code to Prove Software Safety

Anthropic announced on Thursday that its Claude model has completed the first large-scale mathematical proof that existing software code behaves correctly. The company stated that the AI generated approximately 180 million lines of proof code in a matter of weeks, a task that historically required extensive human expertise. This development aims to reduce the cost and time associated with formal verification, a process that ensures software runs without error under specific conditions.

Formal verification has traditionally been too expensive for widespread adoption. Specialists previously required 20 person-years to verify seL4, a small operating system kernel. Consequently, the technique was largely limited to high-stakes environments such as Airbus avionics. Anthropic’s Frontier Red Team partnered with Theorem to automate this process. The system used test programs from 123 open-source libraries in Google’s OSS-Fuzz project as a foundation for the mathematical analysis.

Claude did not write new software. Instead, it converted crash tests into mathematical problems and analyzed compiled machine code to construct proofs. A separate tool named Rocq automatically checks these proofs. Anthropic described the proof generation as fully autonomous, though human engineers at Theorem built the underlying formal tools. During the process, Claude identified hundreds of bugs and was authorized to fix them. Researchers noted that many of these issues would likely have escaped standard testing because they appear only under narrow conditions, such as unusually large input files.

The testing yielded immediate results. In one week, the team found as many security bugs in a widely used PDF reader as are usually reported for that software in a full year. In another instance, 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 often reveals real defects.

The company emphasized that formally verified does not mean error-free. The proofs hold only under defined constraints, including a specific processor architecture, single-threaded programs, and inputs smaller than 64 MB. Cost remains a factor, as individual proofs consume hundreds of millions of tokens. Anthropic has not published total project costs but argues that reusing proofs can lower expenses. Data from a smaller test run showed that Opus 5.5 solves half the proofs using about 40 percent fewer tokens than Opus 5.

The project shifts the focus from whether AI can program to whether it can prove the safety of that code. Anthropic expects formal verification with AI to become as practical as standard testing. Theorem’s roadmap targets critical infrastructure components, including glibc, OpenSSL, and cURL, for future verification efforts.