Claude writes security proofs for 123 open-source libraries

Anthropic and the start-up Theorem pointed Claude at more than 100 open-source libraries and had it prove security properties on their compiled RISC-V code, the instructions the chip actually runs, Anthropic wrote. Along the way Claude found bugs that years of automated testing had missed.

The Decoder puts the count at 123 libraries from Google’s OSS-Fuzz security project, with roughly 180 million lines of proof code written in a few weeks, and Claude patched bugs it hit on the way. Each proof passes through Rocq, a proof checker, so Anthropic says people barely need to look. The tools and formal groundwork came from people at Theorem.

Claude Pro, Anthropic’s paid plan for people, costs $20 a month on our price sheet, read 23 October 2026; this proof work ran in Anthropic’s research team with Theorem’s tooling.

Because Rocq accepts or rejects every proof on its own, the result rests on that checker and Theorem’s tooling instead of on Claude’s word, which sets this work apart from AI-written code that people still check by hand.