Claude generates 180 million lines of proof code to verify open-source software

The era of AI simply writing code may be giving way to a new standard where machines mathematically guarantee their work is safe.

In a significant development for software reliability, Anthropic and its partner Theorem have demonstrated that Claude can automatically generate formal proofs for existing software. The model produced approximately 180 million lines of proof code in a matter of weeks, covering 123 open-source libraries from Google's OSS-Fuzz project. This volume of automated verification work previously required immense human effort; for context, a team of specialists needed 20 person-years to formally verify the seL4 operating system kernel. By translating crash tests into mathematical problems and analyzing compiled machine code, the system bypasses the traditional need for extensive manual review.

The verification process relied on an independent program called Rocq to check the generated proofs. Anthropic describes the workflow as "fully autonomous," although the underlying tools and formal foundations were engineered by human experts at Theorem. The model did not write new application code but rather validated existing software. During this process, Claude detected hundreds of bugs that conventional testing methods had failed to catch. These errors often required specific, unusual conditions to manifest, such as large input files or particular access sequences, making them difficult for standard test suites to identify.

The impact of these findings was most visible in the testing of a widely used PDF reader. In a single week, the team discovered as many security vulnerabilities as are typically reported for that software in an entire year. In another instance, Claude identified a logical error that had remained undetected for more than a decade. The researchers note that even when a complete proof cannot be generated, the attempt often surfaces real-world bugs, providing value beyond the final certification.

Despite the progress, Anthropic emphasizes that "formally verified" does not equate to "error-free" in all scenarios. The proofs are valid only under strict constraints, including a specific processor architecture, single-threaded execution, and inputs smaller than 64 MB. The computational cost is also significant, with individual proofs consuming hundreds of millions of tokens, though Anthropic has not disclosed the total project cost. Efficiency is improving, however; newer models like Opus 5.5 solve half the required proofs with 40 percent fewer tokens compared to Opus 5.

This advancement suggests that AI-generated software can be trusted more readily, potentially reducing the time developers spend on manual review. The project implies that the industry question is shifting from whether AI can program to whether it can prove its code is safe. Anthropic anticipates that AI-driven formal verification will soon become as routine as standard testing, while Theorem’s roadmap already targets critical infrastructure components like glibc, OpenSSL, and cURL.

For developers and engineers, this shift could lower costs across the entire coding lifecycle. By allowing AI to verify its own output under defined conditions, the sector may move toward a future where mathematical certainty is a standard feature of software development rather than a luxury reserved for critical systems like avionics.