180 Million Lines: How Anthropic's AI Automated Software Safety Checks

Anthropic has fundamentally altered the landscape of software assurance by deploying its Claude model to mathematically verify existing code. In a collaborative effort with partner Theorem, the company generated approximately 180 million lines of proof code in a matter of weeks. This initiative did not focus on creating new software but rather on rigorously checking the correctness of 123 open-source libraries from Google's OSS-Fuzz project. The move signals a potential leap in software automation, aiming to make formal verification as routine as standard testing.

The need for this technology stems from the historical cost of proving software reliability. Traditionally, formal verification required specialized teams and immense time. For example, verifying the seL4 operating system kernel took 20 person-years, a cost that limited the method to critical systems like Airbus avionics. By leveraging AI, Anthropic's Frontier Red Team and Theorem shifted the workload to Claude. The model translated crash tests into mathematical problems, analyzed compiled machine code, and produced extensive proofs. These outputs were then checked by an independent program called Rocq, a process described as fully autonomous despite the human-built foundations provided by Theorem.

The immediate practical impact was the discovery of numerous flaws that conventional testing had overlooked. Claude identified hundreds of bugs that only appeared under specific conditions, such as unusually large input files or precise access sequences. In one notable instance involving a widely used PDF reader, the team found as many security bugs in a single week as are typically reported in an entire year. In another case, the AI uncovered a logical error that had remained undetected for over a decade. Anthropic notes that even when a complete proof fails, the attempt still surfaces these real-world issues.

However, the term formally verified does not imply that the software is error-free in all contexts. The proofs hold only under strict conditions, including specific processor architectures, single-threaded programs, and inputs under 64 MB. Since these libraries often run on varied hardware and configurations, these limits are significant. The compute requirements are also steep, with individual proofs consuming hundreds of millions of tokens. Anthropic has not disclosed the total project costs, but the efficiency gains are evident in newer models. Opus 5.5 solves half the proofs with roughly 40 percent fewer tokens than its predecessor, Opus 5.

Looking ahead, the success of this project strengthens the argument for deeper software automation. The central question in the industry is shifting from whether AI can program to whether it can prove that its code is safe. Anthropic expects AI-driven formal verification to become as practical as standard testing in the near future. Theorem has already outlined a roadmap targeting critical infrastructure components such as glibc, OpenSSL, and cURL. If these efforts succeed, developers will be able to trust AI-generated software more, reducing the time and cost spent on manual review and verification.

This development marks a significant step toward reducing the friction in software development. By automating the proof process, companies can allocate resources more efficiently while maintaining high safety standards. The collaboration between Anthropic and Theorem demonstrates that AI can handle complex mathematical tasks that were previously reserved for human specialists. As the technology matures and costs decrease, the integration of formal verification into the standard development pipeline appears increasingly likely, potentially transforming how critical software is built and maintained.