Reposted by beat
It's very exciting that AI can now take substantial real-world software like zlib and prove it correct (github.com/kim-em/lean-...).
@kirancodes.me found a bug by fuzzing, but it was outside of the boundary of what was verified (kirancodes.me/posts/log-wh...). That doesn't diminish the excitement!
kirancodes.me
Lean proved this program was correct; then I found a bug.
AI agents are getting very good at finding vulnerabilities in large-scale software systems.