AIs are Getting Better at Finding and Exploiting Internet Vulnerabilities

Anthropic reports that its Claude Sonnet 4.5 model can autonomously identify and exploit public CVEs in high-fidelity simulations, execute multistage attacks across dozens of hosts, and exfiltrate simulated Equifax breach data using only standard open-source tools; the post warns this capability lowers barriers to automated exploitation and stresses the importance of prompt patching and security fundamentals.