Anthropic’s Claude Automates Formal Verification, Uncovering Deep-Seated Software Flaws

Anthropic has utilized its Claude model to mathematically verify existing software, effectively automating a process that previously required years of specialized human labor. Working alongside partner Theorem, the company’s Frontier Red Team generated approximately 180 million lines of proof code for 123 open-source libraries over a few weeks. This initiative addresses a critical gap in AI-generated software by ensuring code correctness through formal proofs rather than traditional testing alone.

The adoption of formal verification has historically been limited by its high cost and complexity. For instance, a team of specialists required 20 person-years to verify the seL4 operating system kernel, a time investment that restricted the method to high-stakes environments like Airbus avionics. To overcome these barriers, Claude translated crash tests from Google’s OSS-Fuzz security project into mathematical problems. The model analyzed compiled machine code and produced extensive proofs, which were then independently checked by a program called Rocq. Anthropic describes this workflow as "fully autonomous," noting that it requires minimal human review despite the foundational tools being developed by Theorem.

A significant outcome of this effort was the discovery of hundreds of bugs that conventional testing had overlooked. These defects typically manifest only under specific conditions, such as unusually large input files or particular access sequences. In one notable instance, the team identified as many security vulnerabilities in a widely used PDF reader in a single week as are typically reported for that software in an entire year. Additionally, Claude detected a logical error in another application that had remained undetected for more than a decade. According to Anthropic, even when a complete proof fails, the attempt still surfaces these latent issues.

Despite the advancements, the company emphasizes that "formally verified" is not synonymous with "error-free." The proofs remain valid only under strict parameters, including a specific processor architecture, single-threaded execution, and inputs smaller than 64 MB. Since real-world libraries often run on diverse hardware and configurations, these limitations represent a practical boundary for the current methodology. Furthermore, the computational cost is substantial, with individual proofs consuming hundreds of millions of tokens. While Anthropic has not disclosed the total project cost, the process demonstrates that AI can verify its own code, potentially reducing the time developers spend on manual review.

The project also highlights improvements in model efficiency. A smaller test run indicated that proofs are reusable and that newer models are more efficient; Opus 5.5 solved half the proofs using approximately 40 percent fewer tokens than its predecessor, Opus 5. If AI can write code and guarantee its correctness under defined conditions, the industry focus may shift from whether AI can program to whether it can prove its safety. This transition could have consequences that extend well beyond current coding tools.

Looking ahead, Anthropic expects formal verification with AI to become as routine as standard testing. Theorem’s roadmap already targets critical infrastructure components, including glibc, OpenSSL, and cURL. As the cost and complexity of verification decrease, developers may place greater trust in AI-generated software, streamlining the entire coding and review process.