Anthropic uses Claude to formally verify existing code under strict conditions

What happened

Anthropic and Theorem used Claude to verify existing software, producing roughly 180 million lines of mathematical proof code in a few weeks from test programs across 123 open-source libraries in Google's OSS-Fuzz project. Rocq independently checked the proofs, and Anthropic says the process uncovered hundreds of bugs with almost no human review. The guarantees apply only to a specific processor architecture, single-threaded programs, and inputs under 64 MB, while individual proofs consumed hundreds of millions of tokens.

Read next

• More changes in today’s briefingOlmoEarth v1.3 releases global satellite embeddings and a browser-based mapping explorer

The story so far

Anthropic and Theorem used Claude to formally verify existing software, generating roughly 180 million lines of mathematical proof code in a few weeks from test programs across 123 open-source libraries in Google’s OSS-Fuzz project. Rocq independently checked the proofs, and Anthropic said the process uncovered hundreds of bugs with almost no human review. The guarantees apply only to a specific processor architecture, single-threaded programs, and inputs under 64 MB. Individual proofs consumed hundreds of millions of tokens, making the computational cost and narrow scope central limits of the work.