CrowdStrike's new AI kill switch is really an agent access cutoff button

Share
Emergency stop switch on an escalator.
Image by Obi on Unsplash (cropped)

As fears spread that unfettered AI could cause systemic damage, the concept of an AI kill switch has emerged. This is the idea, that you hit the big red button (metaphorically at least) and it shuts down the system.

Such an uber button is probably not realistic, but that didn't stop CrowdStrike from announcing an AI kill switch last week at Fal.Con Europe in Barcelona, and making the initial functionality available immediately.

A more apt name, however, might be a rogue agent kill switch. Let's suppose you observe an agent doing something it's not supposed to do such as, say, I don't know, just riffing here, hacking Hugging Face, you could implement this kill switch and cut off access to the underlying model, effectively stopping the agent.

As CrowdStrike describes it, "If an agent breaks out of its intended behavior, a defender can terminate it where it runs, block it from launching again, and cut off its enterprise network access."

For now, that involves blocking an agent's access to the model through CrowdStrike's AI gateway or a third-party gateway. Future capabilities will allow broader agent containment across CrowdStrike's Falcon platform.

It might not stop the underlying models from doing untoward acts, but at least it can put down a misbehaving agent, and that's a start.

This story first ran in FastForward #85: Dispatches from the road, my Friday newsletter for enterprise tech leaders.

Want stories like this in your inbox every week? Subscribe to FastForward.

Read more