AI Model Proves Old Software Works

Anthropic has announced that its Claude model successfully generated mathematical proofs verifying the correctness of existing software code. The company states that this achievement reduces the need for extensive manual review, a task that previously required significant expert resources. The work represents a shift in how developers validate software reliability, moving from traditional testing to automated mathematical confirmation.

Formal verification has historically been expensive and slow. Specialized teams required 20 person-years to verify seL4, a small operating system kernel. This cost structure limited the technique to high-stakes environments such as Airbus avionics. Anthropic’s Frontier Red Team partnered with Theorem to apply Claude to this process. Over a period of weeks, the model produced approximately 180 million lines of proof code. The model utilized test programs from 123 open-source libraries within Google’s OSS-Fuzz security project.

During the verification process, Claude analyzed compiled machine code and constructed proofs checked by a separate program called Rocq. The researchers described the generation as fully autonomous, though human experts at Theorem built the underlying tools. The model discovered hundreds of bugs, many of which likely escaped conventional testing due to narrow trigger conditions. In one instance, the team found as many security bugs in a single week as are typically reported for a specific PDF reader in a full year. Claude also identified a logical error that had persisted for more than a decade.

Anthropic emphasized that formal verification does not guarantee software is entirely error-free. The proofs hold only under specific conditions, including a particular processor architecture, single-threaded execution, and inputs smaller than 64 MB. Real-world deployments often differ from these parameters. Cost remains a factor, as individual proofs consume hundreds of millions of tokens. However, the company noted that efficiency is improving, with Opus 5.5 using 40 percent fewer tokens than its predecessor to solve half the proofs.

The project aims to increase trust in AI-generated software by allowing developers to spend less time on manual review. Theorem’s roadmap targets critical infrastructure, including glibc, OpenSSL, and cURL. Anthropic expects this approach to become as practical as standard testing methods in the near future.