Claude generates 180 million lines of math to verify old software
Anthropic and Theorem claim the work cuts verification time from decades to weeks, though limits remain on hardware and file size.
By Marcus Johnson ·
Developers using open-source libraries may soon face a new standard for trust. Anthropic announced that its Claude model has mathematically proven the correctness of existing software code. The process required the AI to generate approximately 180 million lines of proof code over a few weeks. This stands in sharp contrast to traditional methods, which often take years of specialist labor.
Formal verification has historically been expensive. It took experts 20 person-years to verify the seL4 operating system kernel. Because of this cost, the technique was mostly limited to critical systems like Airbus avionics. Anthropic partnered with Theorem to apply a different approach using the Claude model. The team used test programs from 123 open-source libraries in Google's OSS-Fuzz security project as a starting point.
Claude did not write new software. It analyzed compiled machine code from existing libraries. The model turned crash tests into mathematical problems and constructed proofs for them. A separate program called Rocq automatically checks these proofs. Anthropic states the process requires almost no human reviewers. The researchers describe the proof generation as fully autonomous, though they note humans built the underlying tools and foundations at Theorem.
The work uncovered real defects. Claude encountered hundreds of bugs and was allowed to fix them. Researchers say many of these issues would have escaped conventional testing because they only appear under narrow conditions. In one week of verifying a widely used PDF reader, the team found as many security bugs as are usually reported for that software in a full year. In another case, the model identified a logical error that had gone unnoticed for more than a decade.
Anthropic emphasizes that formal verification does not mean the code is error-free. The proofs only hold under specific conditions. These include a particular processor architecture, single-threaded programs, and inputs smaller than 64 MB. Real-world deployments often run on different hardware and configurations. Cost is also a constraint, as individual proofs consume hundreds of millions of tokens. Anthropic has not published the total project costs.
The company argues that if AI can verify its own code, developers can trust AI-generated software more. This could reduce the time spent on manual review. A smaller test run showed that proofs can be reused and that efficiency is improving. Opus 5.5 solves half the proofs using about 40 percent fewer tokens than the previous version. The project shifts the focus from whether AI can program to whether it can prove the code is safe. Theorem's roadmap targets critical infrastructure including glibc, OpenSSL, and cURL.
Source: The Decoder
