Anthropic and Theorem generate 180 million lines of code to verify software safety

Roughly 180 million lines of mathematical proof code were generated in just a few weeks by a new system developed by Anthropic and Theorem. This massive output represents a significant shift in how software engineers verify the safety of code. The system used the company's Claude model to check existing software for errors, marking a major step toward automated quality assurance.

The project focused on 123 open-source libraries from Google's OSS-Fuzz security project. Rather than creating new applications, the AI verified the correctness of code that already exists. The model translated crash tests into mathematical problems and analyzed compiled machine code. It then produced extensive proofs to demonstrate that the software behaves as expected under specific conditions.

Formal verification has historically been too expensive for most developers. A team of specialists previously required 20 person-years to verify the small operating system kernel seL4. Because of this high cost, the method was largely reserved for critical systems like Airbus avionics. The new approach drastically reduces the time required, allowing for much broader application of rigorous safety checks.

An independent verification program called Rocq checks the proofs generated by Claude. Anthropic states that the process requires almost no human reviewers. The researchers describe the system as fully autonomous, although the underlying tools and formal foundations were built by humans at Theorem. This level of automation suggests a future where AI handles both coding and validation tasks.

The system discovered hundreds of bugs during the verification process. Many of these issues would have likely been missed by conventional testing methods. These bugs often only appear under very specific conditions, such as unusually large input files or particular sequences of accesses. Claude was permitted to fix these errors itself once identified.

In one notable instance, the team verified a widely used PDF reader. They found as many security bugs in a single week as are typically reported for that software in an entire year. In another case, the AI caught a logical error that had remained undetected for over a decade. Anthropic notes that even when a complete proof fails, the attempt still surfaces real bugs in the code.

Despite the progress, the method has strict limits. A label of formally verified does not mean the software is completely error-free. The proofs only hold under specific conditions, including a particular processor architecture and single-threaded programs. Inputs must also be under 64 MB. In real-world scenarios, these libraries often run on different hardware and configurations.

The compute costs for this verification are significant, with individual proofs running up to hundreds of millions of tokens. However, the process is becoming more efficient. Newer models like Opus 5.5 solve half the proofs with about 40 percent fewer tokens than the previous version. This efficiency could make the technology more accessible over time.

Anthropic and Theorem believe this technology will change how developers trust AI-generated software. If AI can write code and prove its correctness, the entire coding process becomes more reliable. Theorem's roadmap already targets critical infrastructure such as glibc, OpenSSL, and cURL. This suggests that formal verification with AI will soon become as routine as standard testing.