Anthropic Deploys Claude to Verify Software Code with Mathematical Precision

Formal verification, the rigorous process of proving software operates without error in all scenarios, has historically been an expensive and slow endeavor reserved for high-stakes environments like aerospace. Today, Anthropic announced that its Claude model has successfully performed this task on existing software, marking a significant shift in how developers might ensure code reliability.

The collaboration involved Anthropic's Frontier Red Team and a partner known as Theorem. Over a period of just a few weeks, Claude generated approximately 180 million lines of mathematical proof code. This massive output was derived from test programs sourced from 123 open-source libraries within Google's OSS-Fuzz security project. The model did not write new software but instead analyzed compiled machine code to transform crash tests into solvable mathematical problems.

A key aspect of this workflow is the use of Rocq, a separate verification program that automatically checks the proofs constructed by Claude. Anthropic describes the proof generation process as fully autonomous, requiring minimal human intervention. While the underlying formal foundations and tools were built by humans at Theorem, the actual generation of proofs was handled by the AI. This setup allows for a high volume of verification work to be completed without the extensive manual effort previously required.

The results of the project highlighted the model's ability to uncover deep-seated issues. Claude identified hundreds of bugs during the verification process and was permitted to fix them. Many of these errors likely would have escaped traditional testing because they only manifest under narrow conditions, such as unusually large input files or specific access sequences. 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 over an entire year.

Despite the impressive findings, the company emphasizes that the term formally verified does not imply the software is completely error-free. The proofs are valid only under specific conditions, including a particular processor architecture, single-threaded execution, and inputs smaller than 64 megabytes. Real-world deployments often involve different hardware and configurations, which means the verification scope is limited. Additionally, the computational cost remains a factor, with individual proofs consuming hundreds of millions of tokens.

Anthropic argues that this capability could make AI-generated software more trustworthy, allowing developers to spend less time on manual reviews and reduce overall coding costs. Early tests suggest that proofs can be reused, and efficiency is improving, with newer model versions solving half the proofs using significantly fewer tokens. The researchers view this work as a pivotal step, shifting the focus from whether AI can code to whether it can prove that code is safe. Theorem's roadmap now targets critical infrastructure libraries such as glibc, OpenSSL, and cURL.

This development suggests a future where formal verification becomes as routine and practical as standard testing. By leveraging AI to handle the heavy lifting of mathematical proof, the industry may see a reduction in the barriers to entry for rigorous software validation. The project demonstrates that while the technology is still maturing and has clear limitations, it offers a promising path toward more reliable and efficient software development processes.