AI Proves Code Safety in Weeks

Roughly 180 million lines of mathematical proof code were generated in a matter of weeks by Anthropic's Claude model. This volume of verification work represents a significant shift in how software safety is assessed. Previously, such comprehensive formal verification required extensive human labor. Now, an artificial intelligence system has completed the task with minimal human intervention. The achievement highlights a new capability in AI-driven development.

Historically, proving that software runs without errors under all conditions was prohibitively expensive. For example, a team of specialists spent 20 person-years to formally verify the seL4 operating system kernel. Because of this high cost, the method was typically reserved for critical systems like avionics in aircraft. Anthropic aims to change this dynamic by making formal verification a routine part of the software development lifecycle. The company views this as a necessary step toward more reliable AI-generated code.

The project was a collaboration between Anthropic's Frontier Red Team and the partner company Theorem. Claude did not write new application code from scratch. Instead, it focused on verifying existing software. The model translated crash tests from 123 open-source libraries into mathematical problems. It then analyzed the compiled machine code to produce extensive proofs of correctness. This process utilized test programs from Google's OSS-Fuzz security project.

An independent program called Rocq automatically checks the proofs generated by Claude. According to Anthropic, this workflow requires almost no human reviewers for the verification step itself. The researchers describe the process as fully autonomous proof generation. However, they note that the underlying formal foundations and tools were developed by human experts at Theorem. This distinction is important for understanding the current limits of the technology.

During the verification process, Claude identified hundreds of bugs that conventional testing had missed. These errors often only appear under very specific conditions, such as unusually large input files or specific access sequences. In one notable instance, the team found as many security bugs in a single week for a popular PDF reader as are typically reported for that software in an entire year. In another case, a logical error that had gone undetected for over a decade was caught by the model.

Anthropic acknowledges that the term formally verified does not mean the software is completely error-free. The proofs are valid only under strict conditions. These include a specific processor architecture, single-threaded execution, and input sizes under 64 megabytes. In real-world practice, these libraries often run on different hardware and configurations. Therefore, the verification provides a strong guarantee within a defined scope but does not cover every possible usage scenario.

The computational cost of this process is significant. Individual proofs can consume hundreds of millions of tokens. While Anthropic has not reported the total project costs, the efficiency gains are notable. A smaller test run showed that newer models are more efficient. Opus 5.5, for instance, solves half the proofs using about 40 percent fewer tokens than the previous version, Opus 5. This trend suggests that the economic barrier to entry will continue to lower over time.

The primary benefit of this technology is increased trust in AI-generated software. If developers can verify that AI-written code is correct, they can spend less time on manual review. This saves costs across the entire coding process, not just verification. The project strengthens the argument for broader software automation. The central question shifts from whether AI can program to whether it can prove its own code is safe.

Looking ahead, Anthropic expects formal verification with AI to become as practical as standard testing. Theorem's roadmap already targets critical infrastructure components such as glibc, OpenSSL, and cURL. If these goals are met, the consequences could extend well beyond coding tools. This development marks a major step in the evolution of software reliability and automation.

For readers in the Mid-Atlantic region, this technological shift signals a future where software reliability is less dependent on manual inspection. As these tools become more efficient, the pace of software development may accelerate. Businesses and developers in the region will likely see these changes impact product release cycles and security standards. The integration of formal verification into standard workflows will be a key factor in this transition.