Claude Automates Software Safety Proofs at Scale

Claude can now mathematically prove that software code is safe without human oversight. This new capability allows AI to verify its own logic and identify deep structural flaws that standard testing methods often miss. The shift marks a significant step toward fully automated software assurance.

The scale of this operation is massive. In just a few weeks, the model generated approximately 180 million lines of proof code. It analyzed test programs from 123 different open source libraries found in Google's OSS-Fuzz security project. These results were achieved by Anthropic's Frontier Red Team in partnership with Theorem.

Traditionally, proving software is error free required immense manual effort. For instance, specialists spent 20 person years just to verify the small seL4 operating system kernel. Because of this cost, such verification was usually reserved for critical systems like Airbus avionics. Now, Claude translates crash tests into mathematical problems and analyzes compiled machine code to produce these proofs.

The process relies on an independent verification program called Rocq. This tool automatically checks the generated proofs. According to Anthropic, the workflow requires almost no human review. While the formal foundations and tools were built by humans at Theorem, the actual proof generation is described as fully autonomous.

During this verification process, Claude discovered hundreds of bugs. It was also permitted to fix them itself. Many of these errors only appeared under specific conditions, such as unusually large input files or specific access sequences. In one notable case involving a widely used PDF reader, the team found as many security bugs in one week as are typically reported for that software in an entire year. Claude also caught a logical error that had remained undetected for over a decade.

Despite these successes, the technology has clear limits. Formal verification does not mean the code is perfect in every environment. The proofs hold only under strict conditions, including a specific processor architecture and single threaded programs with inputs under 64 MB. The computational cost is also high, with individual proofs consuming hundreds of millions of tokens. However, newer models are becoming more efficient. Opus 5.5 solves half the proofs with about 40 percent fewer tokens than the previous version.

If AI can write code and guarantee its correctness, developers can trust automated software more. This could reduce costs across the entire coding process. Anthropic expects this type of verification to become as routine as standard testing soon. Theorem is already targeting critical infrastructure libraries like glibc, OpenSSL, and cURL for this treatment.