180 Million Lines: How Claude Just Automated Software Safety Checks

180 million. That is the approximate number of lines of mathematical proof code that Anthropic’s Claude model generated in just a few weeks. This staggering volume of work represents a significant shift in how we think about software reliability, as the AI didn’t just write new code; it mathematically verified that existing code works correctly. This effort, conducted in partnership with the firm Theorem, addresses a long-standing challenge in computer science: proving that software is truly error-free without having to test every possible scenario manually.

Historically, formal verification has been an incredibly expensive and time-consuming process reserved for high-stakes environments like aerospace. For example, a team of specialists required 20 person-years to formally verify the seL4 operating system kernel. Because of these high barriers, the technique was rarely used outside of critical systems like Airbus avionics. By letting Claude handle the heavy lifting, Anthropic and Theorem managed to apply this rigorous standard to a much broader range of software in a fraction of the time. The model processed test programs from 123 different open-source libraries found in Google’s OSS-Fuzz security project.

The process works by translating standard crash tests into mathematical problems. Claude then analyzes the compiled machine code to produce extensive proofs. These proofs are subsequently checked by an independent verification program called Rocq. According to Anthropic, this workflow requires very little human intervention, leading the researchers to describe it as “fully autonomous” proof generation. However, it is important to note that the underlying tools and formal foundations were still built by humans at Theorem, meaning the AI is operating within a structured framework rather than inventing the math from scratch.

One of the most striking outcomes of this project is the discovery of bugs that had evaded detection for years. While generating proofs, Claude identified hundreds of bugs and was permitted to fix them automatically. Many of these issues only surfaced under very specific conditions, such as unusually large input files or particular sequences of memory accesses, which conventional testing often overlooks. In one notable instance, the team found as many security bugs in a widely used PDF reader in a single week as are typically reported for that software over an entire year. In another case, Claude caught a logical error that had remained undetected for more than a decade.

Despite the impressive results, there are practical limits to what this technology can achieve. Anthropic clarifies that “formally verified” does not mean the software is perfect in every context. The proofs are valid only under strict conditions, including a specific processor architecture, single-threaded execution, and inputs smaller than 64 MB. Since these libraries often run on diverse hardware configurations in the real world, the verification is a powerful tool but not a universal guarantee of safety. Additionally, the computational cost is high, with individual proofs requiring hundreds of millions of tokens to process.

The financial implications of this approach are becoming clearer as newer models improve efficiency. A smaller test run indicated that proofs are reusable, and the latest model, Opus 5.5, solves half the required proofs using about 40 percent fewer tokens than its predecessor, Opus 5. Anthropic has not disclosed the total cost of the project, but the efficiency gains suggest that the expense of verification could drop significantly. If developers can trust AI-generated software more, they can spend less time on manual review, potentially saving costs across the entire coding lifecycle.

This project shifts the conversation in software automation from asking if AI can write code to asking if it can prove that code is safe. If the answer is yes, the impact could extend well beyond basic coding tools. Anthropic expects that formal verification with AI will soon become as routine as standard testing. Theorem’s roadmap already targets critical infrastructure components like glibc, OpenSSL, and cURL, suggesting that this technology is moving from experimental to essential in the coming years.