AI’s ‘middle class’ has gotten dramatically better at hacking

Research from XBOW and Anthropic shows that mid-tier and frontier generative AI models are becoming increasingly effective at offensive security tasks: newer models (e.g., GPT‑5.5, Mythos, Opus variants) find and sometimes exploit vulnerabilities more reliably, multi-agent swarms drastically increase discovered vulnerabilities (at high token cost), and cheaper mid-tier models lower the barrier for attackers. The report highlights that live interaction with targets improves success, source-code access is less critical than real-time testing, and while frontier models remain strongest, affordable middle-class models could pose a growing long-term risk to cybersecurity.