AI Verification: Claude Proves 123 Libraries Safe
Anthropic and Theorem used the model to generate 180 million lines of proof, uncovering bugs that evaded standard testing for years.
By Carmen Wojcik ·
Anthropic confirmed on October 23 that its Claude model generated 180 million lines of mathematical proof to verify 123 open-source software libraries.
The project marks a shift from manual auditing to automated proof generation. For decades, formal verification required specialized teams. Verifying the seL4 operating system kernel alone took experts 20 person-years. Because of that cost, the practice remained rare outside high-stakes fields like aviation software. Anthropic’s Frontier Red Team partnered with Theorem to replace human labor with model-driven computation.
The process began with test programs from Google’s OSS-Fuzz security project. Claude analyzed the compiled machine code of the 123 libraries. It converted crash tests into mathematical problems and constructed proofs to show the code behaves correctly. A separate program named Rocq automatically checks these proofs. The team describes the generation process as fully autonomous, though humans at Theorem built the underlying tools and formal foundations.
The work surfaced hundreds of bugs that traditional methods missed. Many flaws only appear under specific conditions, such as large input files or precise access sequences. In one week, the team found as many security bugs in a widely used PDF reader as are typically reported in a full year. In another case, Claude identified a logical error that had persisted for more than a decade.
Anthropic notes that verification does not guarantee software is error-free. The proofs apply only under defined conditions, including a specific processor architecture, single-threaded execution, and inputs smaller than 64 MB. Individual proofs consume hundreds of millions of tokens. The company has not published total project costs. However, a smaller test run showed that Opus 5.5 solves half the proofs using 40 percent fewer tokens than Opus 5.
The team argues that if AI can verify its own code, developers will trust automated software more. This could reduce the time spent on manual review. Theorem’s roadmap now targets critical infrastructure components, including glibc, OpenSSL, and cURL.
The project changes the focus from AI writing code to AI proving code is safe. Anthropic expects this method to become as practical as standard testing.
