Anthropic (@AnthropicAI) on X

Anthropic's Frontier Red Team has published research on using Claude to formally verify software at scale. Formal verification means mathematically proving that a piece of code cannot exhibit certain classes of bugs, rather than just testing it and hoping. The catch has always been cost: the technique has historically worked on small codebases or demanded a lot of scarce expert time. The pitch here is that AI models are changing that math. The concrete result is the headline number. In collaboration with Theorem Labs, the team used Claude to prove security properties across more than one hundred real-world open-source libraries, and they did it over a span of a few weeks. That is a far cry from the handful of projects that formal methods have traditionally touched. Because the libraries are real and widely used, the work is a more meaningful signal than a toy benchmark, though the post is short on the details a skeptic would want, such as exactly which properties were proven, how often Claude's proofs needed human correction, and how the results were checked. For anyone who ships or depends on open-source code, the practical implication is that verified security guarantees may stop being a luxury reserved for high-assurance projects. If the approach holds up, it could let maintainers prove the absence of certain vulnerability classes across large dependency trees. The full research is at anthropic.com/research/formal-verification, and it's the place to check the methodology and the limits before taking the claims at face value.