Anthropic reported using Claude to formally verify security properties in 123 open-source libraries, uncovering bugs missed by extensive testing.

On October 22, 2026, Anthropic reported that Claude, working with software-verification firm Theorem, proved crash-safety properties for 239 test entry points across 123 open-source libraries from Google’s OSS-Fuzz. Over several weeks, Claude translated existing fuzz tests into formal specifications, patched bugs it uncovered, and generated machine-checkable proofs covering every input under 64 MB. The proofs apply to compiled RISC-V binaries and establish normal exit without memory-safety errors or failed assertions, subject to assumptions including single-threaded execution, static linking, a fixed memory budget, and modeled Linux behavior. The work produced about 180 million lines of Rocq proof code; a typical proof was roughly 600,000 lines. It exposed hundreds of bugs, including a decompressor counter underflow enabling an out-of-bounds write on a rare, roughly 20 MB input.