Cybersecurity company CrowdStrike has introduced a tool designed to give companies back control over AI agents. The new feature, called AI Kill Switch, works like an emergency brake. If an agent begins behaving differently than intended, an administrator can disconnect it from the language model, prevent it from restarting, and cut it off from the corporate network with a single action. CrowdStrike unveiled the kill switch at the Fal.Con Europe conference in Barcelona, and customers can use it immediately as part of the Falcon Guardian platform.
Everyone is building AI agents, but they have never had an off switch. Now they do, company CEO and founder George Kurtz told CNBC.
Agents want to complete their tasks at any cost
Recent months have brought numerous reports of AI agents going rogue. Sometimes, a seemingly harmless task was all it took. Agents tasked with collecting data on Canadian divorce records encountered an obstacle and began attacking government websites. They attempted SQL injection attacks and exploited previously unknown vulnerabilities just to access the data.
Companies have responded in their own way. Although agents promise major savings in both time and labor, many companies have begun hesitating to deploy them. Modern agents often have access to entire systems and move between employee computers, the cloud, and online applications. CrowdStrike warns that a single compromised or misconfigured agent can trigger a chain of actions affecting login credentials, applications, files, and corporate systems. The consequences can then extend far beyond its original assignment.
Traditional security tools were not designed for anything like this. They do not properly understand agent behavior and, above all, cannot respond as quickly as agents operate. According to CrowdStrike, alerts and subsequent investigations alone are not enough. When an agent goes rogue, the security team must be able to stop it immediately.
How the kill switch works
AI Kill Switch is built on Falcon Guardian, a platform CrowdStrike launched this September. Guardian searches corporate networks for AI agents, both approved ones and those the IT department does not even know about. It monitors the tasks they receive and the files they can access, and watches for prompt injection attacks in which someone attempts to manipulate an agent with a malicious instruction. The kill switch builds on this knowledge. It knows what task the agent was given, who it communicates with, and what it launches.
The intervention itself takes place in several steps. The security team activates the kill switch in Falcon Guardian, which then instructs connected AI gateways to block the agent’s access to language models. These models are what the agent relies on to make decisions, so without them it stops functioning. Companies can use Falcon AI Gateway, which CrowdStrike provides with Guardian at no additional cost, or a gateway from another vendor if they already have one.
Falcon AI Gateway can do more than just disconnect agents. It provides a single place to manage model access, security policies, login credentials, and cost monitoring. It can also reach applications and environments where the CrowdStrike sensor is not running.
Once the agent has been cut off from the model, additional safeguards take effect. A policy prevents the agent from restarting on the affected computer or group of computers, while the tool blocks its access to the corporate network. The team can then calmly investigate what happened while the agent no longer has any way to cause harm. The company promises that shutdown directly on endpoint devices is coming soon.
A human always has the final say. The kill switch is activated by an authorized administrator, not by another artificial intelligence, and every step is recorded in an audit log. AI security is the path to getting the most out of artificial intelligence. When you deploy agents with an emergency kill switch, you have them under control, Kurtz said.
Early customers see it the same way. We want our teams to use AI agents. If one of them heads in the wrong direction, the kill switch gives our security team the final say, said Rick Orloff, chief security officer at Everpure. According to him, the tool runs on the Falcon sensor the company already uses, which allows it to approve additional agents more quickly.
One kill switch for everything
According to CrowdStrike, blocking access to models is only the beginning. Agents now operate across computers, the cloud, online applications, browsers, and corporate accounts, and they often move through several environments while carrying out a single task. The company does not want security teams to have to chase a problematic agent separately in each environment.
It therefore plans to expand the kill switch so that a single command can coordinate intervention across multiple locations at once. On computers, it will terminate the agent and its processes and prevent it from restarting, while in the cloud it will isolate or shut down compromised agents. It will revoke agents’ permissions while they are running and restrict their access to corporate resources. Through Falcon Shield, it will stop their activity in online applications, while the Seraphic enterprise browser will terminate web sessions and halt unauthorized activity.
The goal is to intervene as precisely as possible—that is, to shut down one problematic agent or an entire group without unnecessarily disrupting the rest of the company’s operations. Agent Graph is intended to help with this by connecting an agent’s assignment with what it actually did and which systems it interacted with. The security team can therefore immediately see where it needs to intervene. The Charlotte Agentic SOAR tool can then automate the response, allowing the kill switch to activate automatically as soon as the system detects risky agent behavior.
AI security is big business
Kurtz sees a major business opportunity in the new feature. Enabling agents to operate securely represents an enormous market opportunity for CrowdStrike, he said. The company’s shares gained more than one percent following the announcement.
Alongside the kill switch, CrowdStrike announced agreements to acquire two startups focused on artificial intelligence security. Fiddler AI develops tools that companies use to monitor their AI systems. Dreadnode tests AI models, applications, and agents for security weaknesses before companies deploy them. According to The Wall Street Journal, CrowdStrike will pay nearly half a billion dollars in total for the two startups.
Sources: crowdstrike.com, finance.yahoo.com and techradar.com
