Anthropic and Theorem Use Claude to Automate Formal Software Verification

Anthropic has demonstrated that its Claude model can automatically generate mathematical proofs to verify the correctness of existing software, a capability that could fundamentally alter how developers ensure code safety. Working with partner Theorem, the company utilized the AI to perform formal verification, a process that traditionally requires extensive manual labor from specialized mathematicians. This recent advancement marks a shift from AI merely writing code to AI proving that the code behaves as intended under specific conditions.

The project focused on verifying 123 open-source libraries from Google’s OSS-Fuzz security initiative. Over the course of a few weeks, Claude generated approximately 180 million lines of proof code. Unlike traditional approaches where humans write new code, this process involved translating existing crash tests into mathematical problems and analyzing compiled machine code. The resulting proofs were checked by an independent verification program called Rocq, with Anthropic noting that the workflow required minimal human intervention, effectively operating as a fully autonomous system.

A significant outcome of the verification process was the discovery of hundreds of bugs that had evaded conventional testing for years. These defects typically only manifest under highly specific circumstances, such as unusual input file sizes or precise sequences of memory access. In one notable instance involving a widely used PDF reader, the team identified as many security vulnerabilities in a single week as are typically reported for that software over an entire year. Claude also detected a logical error in another application that had remained undetected for more than a decade, demonstrating the depth of its analytical reach.

Despite these successes, Anthropic emphasizes that formal verification does not equate to absolute error-freedom. The proofs are valid only under strict constraints, including specific processor architectures, single-threaded execution, and input sizes under 64 megabytes. Since real-world deployments often involve different hardware configurations, these limitations remain relevant. Additionally, the computational cost is substantial, with individual proofs consuming hundreds of millions of tokens, though the company has not disclosed the total financial cost of the project.

The efficiency of the process is improving rapidly. Data indicates that newer model versions are becoming more token-efficient; Opus 5.5, for example, solves half the proofs using approximately 40 percent fewer tokens than its predecessor, Opus 5. This efficiency gain, combined with the reusability of generated proofs, suggests that the long-term cost of verification will decrease. If AI can reliably verify its own output, developers may spend less time on manual code review, reducing costs across the entire software development lifecycle.

The initiative positions formal verification as a routine step in software development, similar to standard testing. Theorem’s roadmap already targets critical infrastructure components such as glibc, OpenSSL, and cURL. As the technology matures, the focus of the industry may shift from whether AI can program to whether it can guarantee the safety of its creations, potentially transforming the automation of critical software systems.

This development addresses a core challenge in AI-generated software: trust. By providing mathematical guarantees of correctness, the collaboration aims to bridge the gap between code generation and code reliability. While practical limits exist, the ability to automatically surface deep-seated bugs and produce rigorous proofs suggests that the future of software engineering may rely heavily on AI-driven verification tools.