Claude produced machine-checked memory-safety proofs for 123 open-source libraries
Anthropic's red team and the startup Theorem had Claude write Rocq-verified proofs that fuzzing entry points in 123 OSS-Fuzz libraries are free of memory-safety errors, finding and patching hundreds of bugs on the way. The guarantees are narrow: single-threaded RISC-V builds with bounded input sizes.
Why it matters
Suggests LLM-driven formal verification of real code is becoming practical, not just fuzzing.
Source: Read the full story on Unite.AI