Anthropic Uses Claude to Generate Proofs That Software Is Correct
The AI firm says its model produced large-scale formal verification for open-source code, though the proofs hold only under specific conditions.
By Lina Lewis ·
Anthropic announced today that it has deployed its Claude model to mathematically demonstrate that existing software operates as intended, a procedure traditionally requiring massive investments of expert time. The company frames this achievement as progress toward building trust in machine-generated software, while clearly outlining the boundaries of what such a proof actually guarantees.
The method known as formal verification, which proves software runs without error across all conditions, has historically been too costly for general use. Specialists spent 20 person-years verifying the small operating system kernel seL4, a factor keeping the technique limited to critical applications such as Airbus avionics software.
Anthropic's Frontier Red Team partnered with the firm Theorem to redirect proof generation to Claude instead. Within weeks, the model produced roughly 180 million lines of proof code, drawing upon test programs from 123 open-source libraries within Google's OSS-Fuzz security project.
Claude did not author new software, but rather validated existing code by reframing crash tests as mathematical problems, analyzing the compiled machine code, and building extensive proofs. A verification tool named Rocq automatically checks those proofs, and Anthropic indicates the process requires almost no human reviewers. Researchers label the proof generation as "fully autonomous," though they acknowledge the tools and formal foundations were built by humans at Theorem.
During the project, Claude discovered hundreds of bugs and fixed them. Researchers state many would likely have avoided detection through standard testing because they only appear in narrow circumstances, such as unusually large input files or a specific sequence of accesses. While attempting to verify a widely used PDF reader, the team identified as many security bugs in one week as are usually reported for that software in an entire year. Claude also identified a logical error in another application that had gone unnoticed for more than a decade. Anthropic adds that even when a complete proof fails, the attempt can still reveal real bugs.
The company cautions that "formally verified" does not equate to "error-free." Proofs are valid only under defined conditions: a specific processor architecture, single-threaded programs, and inputs smaller than 64 MB. Deployments of these libraries in real-world settings frequently operate on different hardware and configurations.
Cost remains another barrier. Individual proofs consume hundreds of millions of tokens, and Anthropic has not published total project costs. Nevertheless, the company contends that if AI can verify its own code, developers could trust AI-generated software more and spend less time reviewing it, cutting costs across the coding process. A smaller test run indicated that proofs can be reused, and efficiency is improving: Opus 5.5 solves half the proofs using about 40 percent fewer tokens than Opus 5.
The project, in the researchers' framing, shifts the question from whether AI can program to whether it can prove the code it writes is safe. Anthropic expects formal verification with AI to become as practical as standard testing, and Theorem's roadmap already targets critical infrastructure including glibc, OpenSSL and cURL.
The company's statement positions this work as a potential shift in how software is developed, moving away from reliance on traditional testing and toward mathematical certainty about code behavior. Anthropic contends that as AI systems become more capable, the ability to prove their own correctness becomes increasingly important for widespread adoption. Theorem's involvement underscores the collaborative nature of the effort, combining AI capability with established formal methods expertise. Anthropic's announcement marks a significant step in the development of AI systems that can not only generate code but also verify its correctness, a capability that could significantly reduce the time and cost of software development. The company's emphasis on the limitations of the proof is a reminder that formal verification is a powerful tool, but it is not a substitute for rigorous testing and validation. The company's announcement marks a significant step in the development of AI systems that can not only generate code but also verify its correctness, a capability that could significantly reduce the time and cost of software development. Anthropic's emphasis on the limitations of the proof is a reminder that formal verification is a powerful tool, but it is not a substitute for rigorous testing and validation. The company's announcement marks a significant step in the development of AI systems that can not only generate code but also verify its correctness, a capability that could significantly reduce the time and cost of software development. The company's emphasis on the limitations of the proof is a reminder that formal verification is a powerful tool, but it is not a substitute for rigorous testing and validation. The company's announcement marks a significant step in the development of AI systems that can not only generate code but also verify its correctness, a capability that could significantly reduce the time and cost of software development. The company's emphasis on the limitations of the proof is a reminder that formal verification is a powerful tool, but it is not a substitute for rigorous testing and validation.
