Anthropic’s Claude verifies software safety with millions of lines of proof

Anthropic has shown that its Claude model can mathematically prove that existing software is correct without human intervention. This new approach addresses a major hurdle in artificial intelligence by ensuring code reliability through formal verification.

Proving that software runs without errors under all conditions has historically been too expensive for most projects. A small operating system kernel once required twenty person-years of specialist work to verify. Because of this cost, such rigorous checks were usually reserved for critical systems like airplane avionics. Now, Anthropic’s Frontier Red Team and partner Theorem have changed that dynamic. They allowed Claude to handle the proof generation work directly.

The AI processed test programs from 123 open-source libraries in Google’s OSS-Fuzz security project. In just a few weeks, Claude produced roughly 180 million lines of mathematical proof code. The model did not write new software. Instead, it translated crash tests into math problems and analyzed compiled machine code. An independent tool called Rocq automatically checks these proofs. According to Anthropic, the process requires almost no human reviewers, a level of automation the researchers describe as fully autonomous.

The verification process also revealed significant defects. Claude encountered hundreds of bugs and was permitted to fix them itself. Many of these issues would have escaped standard testing because they only appear under specific conditions, such as unusually large files or rare data sequences. In one instance, the team found as many security bugs in a popular PDF reader in a single week as are typically reported for that software in an entire year. In another case, the model caught a logical error that had remained undetected for more than ten years.

There are limits to what this method can guarantee. Anthropic notes that a formal verification does not mean the software is perfect in every scenario. The proofs only hold under strict conditions, including specific processor architectures and single-threaded programs with inputs under 64 MB. The computational cost is also high, with individual proofs using hundreds of millions of tokens. However, a smaller test run showed that newer models are becoming more efficient. Opus 5.5 solved half the proofs using about 40 percent fewer tokens than its predecessor, Opus 5.

If AI can write code and guarantee its correctness, the industry may see a major leap in automation. Developers could trust AI-generated software more and spend less time on manual reviews. This shift would save costs across the entire coding process. Theorem’s roadmap already targets critical infrastructure components like glibc, OpenSSL, and cURL. Anthropic expects this type of AI-driven verification to become as routine as standard testing in the near future.