180 Million Lines of Proof: Claude Automates Software Safety Checks

Anthropic’s Claude model has generated approximately 180 million lines of mathematical proof code to verify the correctness of existing software. This massive output was produced in just a few weeks, marking a significant shift in how developers approach software safety. The work involved analyzing 123 open-source libraries and translating complex crash tests into formal mathematical problems. This effort demonstrates a new capability for AI to not just write code, but to rigorously prove that specific code behaves as intended under defined conditions.

The process required a partnership between Anthropic and the company Theorem. The model did not create new software from scratch. Instead, it took existing programs and subjected them to strict logical verification. Claude analyzed the compiled machine code and translated test scenarios into mathematical statements. An independent verification program called Rocq then checked the validity of these proofs. According to Anthropic, the system operated with almost no human review, a state the researchers describe as fully autonomous proof generation.

One of the most striking outcomes was the discovery of hundreds of bugs that conventional testing had failed to catch. These errors often resided in obscure corners of the code, triggered only by very specific conditions such as unusually large input files or particular sequences of data access. In one notable case involving a widely used PDF reader, the team found as many security vulnerabilities in a single week as are typically reported for that software in an entire year. In another instance, the model identified a logical error that had remained undetected for more than a decade.

The efficiency gains compared to traditional methods are substantial. Historically, formal verification was prohibitively expensive and time-consuming. For example, a team of specialists required 20 person-years to verify the small operating system kernel seL4. Because of this cost, the method was usually reserved for critical systems like Airbus avionics. By automating the proof generation, Anthropic and Theorem have drastically reduced the time required to perform these checks, making the process feasible for a broader range of software projects.

Despite the breakthrough, the method has strict practical limitations. Anthropic clarifies that the label formally verified does not mean the software is completely error-free in all possible scenarios. The proofs only hold under specific constraints, including a particular processor architecture, single-threaded execution, and input sizes under 64 MB. In real-world deployments, these libraries often run on diverse hardware and in more complex configurations. Therefore, the verification provides a strong guarantee within a defined scope rather than a universal one.

The computational cost of the project is also a factor to consider. Individual proofs required hundreds of millions of tokens to generate. While Anthropic has not disclosed the total project costs, the company notes that newer models are becoming more efficient. For instance, the Opus 5.5 model can solve half of the proofs using about 40 percent fewer tokens than its predecessor, Opus 5. This improvement suggests that the financial barrier to entry for AI-driven verification may continue to drop over time.

The project suggests that formal verification with AI could become as routine as standard testing in the near future. Theorem’s roadmap already targets critical infrastructure components such as glibc, OpenSSL, and cURL. If AI can reliably prove the safety of code it writes or verifies, developers can place greater trust in automated software processes. This shift could reduce the time spent on manual code review and lower the overall costs of the development lifecycle.

The broader implication is a change in the fundamental question for software engineering. The industry is moving from asking if AI can program to asking if it can prove its code is safe. If the answer remains yes, the potential impact extends well beyond coding tools. It could transform how critical software is built and maintained across various industries, from aviation to finance, by providing a mathematical guarantee of correctness within specific boundaries.