AI Proves Legacy Code Safety

Anthropic states that its Claude model generated approximately 180 million lines of mathematical proof code to verify the correctness of existing software. The company collaborated with partner Theorem to apply formal verification to 123 open-source libraries from Google's OSS-Fuzz project. This work completed within weeks, a timeframe that contrasts sharply with traditional methods.

Formal verification has historically been a costly endeavor due to the intensive labor required. Specialists previously needed 20 person-years to verify seL4, a small operating system kernel. Because of this expense, the technique remained limited to high-stakes sectors like Airbus avionics. Anthropic's Frontier Red Team shifted this process to an automated system.

The system operated by converting crash tests into mathematical problems. Claude analyzed compiled machine code to construct these proofs without writing new software. A separate verification program named Rocq automatically checked the generated proofs. The researchers described the generation process as fully autonomous, although humans at Theorem built the underlying tools and formal foundations.

During the verification run, Claude identified hundreds of bugs and was authorized to fix them. Many of these defects likely would have escaped conventional testing because they only appear under narrow conditions. These conditions include unusually large input files or specific sequences of accesses. In one case involving a widely used PDF reader, the team found as many security bugs in one week as are typically reported for that software in a full year.

Claude also located a logical error that had remained unnoticed for over a decade. The researchers noted that even failed proof attempts can reveal real bugs. However, the company clarified that formal verification does not guarantee error-free status. The proofs hold only under defined conditions, including specific processor architectures, single-threaded programs, and inputs smaller than 64 MB.

Cost remains a factor in this approach. Individual proofs consume hundreds of millions of tokens, though Anthropic has not released total project costs. The company argues that if AI can verify its own code, developers can trust AI-generated software more and reduce review time. 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 AI programming to AI proving code safety. Anthropic expects this method to become as practical as standard testing. Theorem's roadmap already targets critical infrastructure, including glibc, OpenSSL, and cURL.