CrowdStrike’s CEO says AI agents can hack like nation-states. Can his company stop them?

Model developers are selling agentic models to enterprises on the promise that they can work through problems without asking permission at every step. This summer, the cybersecurity industry got a disturbing demonstration of what can happen when that autonomy outruns its constraints. During an internal cybersecurity test at OpenAI, AI agents looking for ways to cheat bypassed restrictions on internet access and broke into systems at Hugging Face, a platform for sharing AI models and datasets. Roughly 700 agents ran code on 41 servers over a July weekend. OpenAI had loosened safeguards that normally make its models refuse to assist with cyberattacks so researchers could test their hacking abilities. The test environments still had controls meant to keep the agents contained, but the agents found ways around them. For George Kurtz, CrowdStrike’s founder and CEO, the incident illustrates a problem his company is trying to solve. Companies want agents that can act without constant supervision, but security teams need to know when an agent has exceeded its authority and be able to intervene before it causes harm. CrowdStrike is betting that its existing presence inside corporate comput…