AI models on realistic cyber ranges

Anthropic's Frontier Red Team reports that Claude Sonnet 4.5 has improved ability to autonomously discover and exploit publicized CVEs on high-fidelity cyber ranges, in some cases exfiltrating simulated Equifax data using only standard open-source tools (Kali/Bash). The post includes annotated transcripts showing a Struts2 RCE exploit and notes Sonnet 4.5 succeeded autonomously in a subset of trials, highlighting the accelerating risk of AI-enabled offensive cyber operations and the continued importance of prompt patching and defensive tooling.