Claude Proves 123 Libraries Safe

Anthropic has used its Claude model to generate roughly 180 million lines of mathematical proof code in a few weeks. The effort verified that 123 open-source software libraries run correctly under specific conditions. The company collaborated with partner Theorem on this project, which aimed to reduce the heavy human labor usually required for formal verification.

Traditionally, proving software works without error has been incredibly expensive. It took specialists 20 person-years to verify seL4, a small operating system kernel. Because of this cost, the technique was mostly limited to high-stakes areas like Airbus avionics. Anthropic argues that using AI changes the cost structure, making the process accessible for a wider range of software.

The model did not write new software. Instead, it analyzed existing code by turning crash tests from Google's OSS-Fuzz project into mathematical problems. Claude examined compiled machine code and built proofs that were then checked by a separate program called Rocq. The researchers described the proof generation process as fully autonomous, with minimal human intervention needed for the final checks.

The project uncovered hundreds of bugs in the tested libraries. Many of these errors likely would have escaped standard testing because they only appear under narrow conditions, such as unusually large input files. In one specific case, the team found as many security bugs in a single week for a popular PDF reader as are typically reported for that software in an entire year. Claude also identified a logical error that had gone unnoticed for more than decade.

Anthropic is clear that formal verification does not mean the software is perfect. The proofs only hold under defined conditions, including a specific processor architecture and inputs smaller than 64 MB. Real-world deployments often use different hardware. The company also notes that individual proofs consume hundreds of millions of tokens, though it has not published the total cost for the 123 libraries.

Efficiency is improving as the technology matures. A smaller test run showed that Opus 5.5 solves half the proofs using about 40 percent fewer tokens than its predecessor, Opus 5. Proofs can also be reused, which may lower future costs. Theorem's roadmap now targets critical infrastructure components, including glibc, OpenSSL, and cURL.

The work shifts the focus from whether AI can write code to whether it can prove that code is safe. Anthropic expects formal verification with AI to become as common as standard testing. For now, the 123 verified libraries have a mathematical guarantee of correctness within the specific limits defined by the researchers.