AI Finds 27-Year-Old Bug Humans Missed

Anthropic's Claude Mythos Preview autonomously discovered a critical vulnerability hidden in OpenBSD's TCP stack for 27 years, surviving decades of human audits and automated testing. The bug, exploitable with just two packets to crash any affected server, was found for under $50 in compute costs. The capability leap is dramatic: Mythos achieved a 90x improvement over its predecessor on Firefox exploit writing, scored 77.8% on SWE-bench Pro, and hit 100% on Anthropic's Cybench CTF, forcing evaluators to pivot to real-world zero-day discovery as the only remaining meaningful benchmark.