Post

LLMs are making formal software verification feasible at a much larger scale

Anthropic and Theorem report using Claude to prove security properties in the RISC-V assembly of more than 100 real-world open-source libraries, and to uncover bugs missed despite extensive testing. This is a promising demonstration of AI helping with the labor-intensive work of formal methods—not proof that arbitrary software can now be verified end to end.

The 209-comment HN discussion welcomed the potential but stressed the limits: proofs only establish what the chosen specification says, and the reported work does not cover every behavior or concurrency case. One commenter described a separate Rust/C-to-Lean experiment that has checked about 12% of PostgreSQL code in single-threaded execution; replies debated the missing concurrency and compiler-correctness work. Those caveats matter as much as the capability claim.