Claude Generates 180 Million Lines of Proof to Certify Open-Source Code Safety

Anthropic’s AI model, Claude, has successfully generated approximately 180 million lines of mathematical proof code to verify the correctness of existing software, marking a significant shift in how code safety is established. This development, achieved in collaboration with partner Theorem, allows the model to formally prove that code operates as intended under specific conditions, effectively automating a process that previously required extensive manual labor from specialized teams.

The initiative addressed the high cost and labor intensity of formal verification, a method historically reserved for critical systems like aviation software due to the massive time investment required. By leveraging test programs from 123 open-source libraries within Google’s OSS-Fuzz project, Claude translated crash tests into mathematical problems and analyzed compiled machine code to produce these proofs. An independent verification tool named Rocq automatically checked the outputs, a process described by researchers as "fully autonomous" because it required minimal human intervention during the generation phase.

Beyond creating proofs, the system identified hundreds of software bugs that conventional testing methods had failed to detect. Many of these issues only manifested under extreme conditions, such as unusually large input files or specific access sequences. During the verification of a widely used PDF reader, the team discovered as many security flaws in a single week as are typically reported for that software in an entire year. In another instance, Claude uncovered a logical error that had remained undetected for over a decade, demonstrating that even failed proof attempts can surface critical defects.

Despite the breakthrough, Anthropic emphasizes that "formally verified" does not equate to "error-free." The proofs are valid only under strict parameters, including a specific processor architecture, single-threaded execution, and input sizes under 64 MB. The computational cost is also significant, with individual proofs consuming hundreds of millions of tokens. However, efficiency is improving; newer models like Opus 5.5 require about 40 percent fewer tokens to solve half the proofs compared to its predecessor, Opus 5.

This project suggests that the focus of AI-driven software development is shifting from simple code generation to guaranteed correctness. If AI can both write code and prove its safety, developers may place greater trust in automated systems, reducing the time spent on manual reviews. Theorem’s roadmap now targets critical infrastructure components such as glibc, OpenSSL, and cURL, indicating that formal verification could soon become as routine as standard testing in the software development lifecycle.