Anthropic and Theorem Generate 180 Million Lines of Proof to Verify Open Source Code

Anthropic used its Claude model to mathematically verify existing software code, generating approximately 180 million lines of proof in a matter of weeks. This work, conducted with partner Theorem, represents a shift in how developers might trust AI-generated or legacy software.

Formal verification has historically been expensive, requiring 20 person-years to verify the seL4 operating system kernel. Consequently, the technique was largely limited to high-stakes environments like Airbus avionics. By automating this process, Anthropic aimed to reduce the resource barrier that kept formal verification out of standard development workflows.

Claude processed test programs from 123 open-source libraries in Google's OSS-Fuzz project. The model analyzed compiled machine code and constructed proofs that were automatically checked by a separate program called Rocq. The team described the proof generation as fully autonomous, noting that human reviewers played a minimal role in the verification step itself.

The project uncovered hundreds of bugs that Claude was authorized to fix. Many of these errors appeared only under narrow conditions, such as unusually large input files or specific access sequences. During a one-week verification of a widely used PDF reader, the team found as many security bugs as are typically reported for that software in an entire year. In another instance, Claude identified a logical error that had remained undetected for over a decade.

Anthropic clarified that formal verification does not guarantee software is error-free in all scenarios. The proofs hold only under specific conditions, including a particular processor architecture, single-threaded execution, and inputs smaller than 64 MB. Real-world deployments often run on different hardware and configurations, meaning the verification scope is limited.

Cost remains a factor, as individual proofs consume hundreds of millions of tokens. Anthropic has not published total project costs but noted that efficiency is improving. A smaller test run showed that Opus 5.5 solves half the proofs using about 40 percent fewer tokens than Opus 5. The company argues that reusing proofs and improving efficiency will lower costs across the coding process.

Theorem's roadmap now targets critical infrastructure libraries including glibc, OpenSSL, and cURL. Anthropic expects formal verification with AI to become as practical as standard testing, moving the focus from whether AI can program to whether it can prove its code is safe.