Anthropic and Theorem Use Claude to Autonomously Verify 123 Libraries

Anthropic and its partner Theorem used the Claude model to mathematically verify existing software code in a matter of weeks. The model generated approximately 180 million lines of mathematical proof code to confirm the behavior of 123 open-source libraries.

This work drew on test programs from Google’s OSS-Fuzz security project. Traditionally, formal verification requires immense human effort. Specialists once needed 20 person-years to verify seL4, a small operating system kernel. This high cost limited the technique to rare cases like Airbus avionics software.

The process did not involve writing new software. Claude analyzed compiled machine code from existing libraries and turned crash tests into mathematical problems. The model constructed extensive proofs to validate the logic. A separate verification program called Rocq checked these proofs automatically. Anthropic states the process requires almost no human reviewers.

The researchers describe the proof generation as fully autonomous, though human experts at Theorem built the underlying tools. During the verification, Claude encountered hundreds of bugs and was permitted to fix them. The team notes many of these errors would likely escape conventional testing because they only appear under narrow conditions, such as unusually large input files.

In one specific test of 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, the model identified a logical error that had gone unnoticed for more than a decade. Anthropic adds that even failed proof attempts can reveal real bugs.

The company notes that formal verification does not mean the software is error-free. The proofs hold only under specific conditions, including a particular processor architecture, single-threaded programs, and inputs smaller than 64 MB. Real-world deployments often run on different hardware and configurations.

Individual proofs consume hundreds of millions of tokens, though total project costs remain unpublished. Anthropic argues that if AI can verify its own code, developers can trust AI-generated software more and spend less time reviewing it. A smaller test run showed that Opus 5.5 solves half the proofs using about 40 percent fewer tokens than Opus 5. Theorem’s roadmap targets critical infrastructure including glibc, OpenSSL, and cURL.