Anthropic's Claude autonomously generated formal proofs for existing code and found hundreds of bugs
Anthropic, with partner Theorem and its Frontier Red Team, used the Claude model to translate crash tests into formal mathematical proofs for 123 OSS-Fuzz libraries, producing about 180 million lines of proof code in weeks and surfacing hundreds of bugs; proofs are checked by an independent verifier called Rocq but are limited to strict environments and high compute costs. Anthropic says the process is largely autonomous and sees AI-driven formal verification becoming as routine as standard testing, while Theorem targets critical libraries like glibc, OpenSSL and cURL next.