At Fal.Con Europe 2026 in Barcelona on November 3, 2026, CrowdStrike announced the "AI Kill Switch," a feature that stops AI agent activity, and made it available the same day. According to the company's announcement, what is available now is the ability to cut off connections to models through "AI Gateway," which centrally manages communication with AI. A feature to stop agents running on endpoints themselves is planned for a later date.
CrowdStrike's "Falcon Guardian" has been generally available since September and handles AI detection, monitoring, and policy enforcement. The AI Kill Switch adds a shutdown control to these existing management functions. However, blocking communication through the Gateway and halting execution on an endpoint cover different scopes of control. Keeping that difference in mind also clarifies how the two acquisitions announced the same day are meant to complement the existing features.
Administrators decide when to stop an agent, and restarts are prevented
CrowdStrike explains that AI agents can take unintended actions using the legitimate permissions they have been given, even when no outside attack is under way. Detecting such problems and actually stopping the agent are separate challenges. With the AI Kill Switch, an administrator with appropriate privileges issues the stop command, and the action is recorded in an audit log.
Beyond cutting network connections, CrowdStrike also describes a mechanism that applies policies to target hosts (computers) or host groups to prevent agents from restarting. However, this description includes the endpoint-side functionality that will be delivered later, and preventing restarts is not available on every operating system or configuration. It is necessary to distinguish between what is offered now and what will be added in the future.
CrowdStrike CEO George Kurtz said in the announcement, "If you deploy agents with a kill switch, you stay in control." His remark emphasizes a mechanism that goes beyond monitoring AI behavior, letting administrators intervene when needed and review their actions afterward. That said, no results from independent testing of the shutdown feature's effectiveness have been presented.
What's the difference between cutting connections at the Gateway and halting execution on an endpoint?
AI Gateway, which has now reached general availability, centrally manages communication with AI models and handles inspection of prompts and responses, monitoring of usage, and enforcement of policies. CrowdStrike's technical blog describes the current AI Kill Switch as a mechanism for rapidly cutting off connections to AI models that pass through the Gateway. In other words, what it can directly control at present is model connections routed through the Gateway, which is different from a feature that stops the agent itself on an endpoint. Companies considering adoption need to understand which routes their AI agents use to connect to models.
Falcon Guardian, meanwhile, includes a mechanism that monitors the processes agents launch on endpoints, the commands they run, and the resources they access, and links these to communication content and instructions given to the agent. This is designed to let organizations trace what actions an AI agent took. However, being able to monitor activity on an endpoint is not the same as being able to stop a running agent. The endpoint-side function of the AI Kill Switch, which addresses the latter, is planned for a future release.
The difference is easy to see when you consider an AI model that runs only in a local environment without a network connection. CrowdStrike itself explains that monitoring the network alone cannot reveal the behavior of such local models. Therefore, cutting off model connections via the Gateway does not necessarily stop local processing that requires no communication. In addition, the announcement gave no figures for response time to shutdown or false-stop rate, and it has not confirmed effectiveness in configurations that bypass the Gateway or across all operating systems.
Fiddler AI and Dreadnode acquisitions to strengthen tracking and testing
Operating AI agents safely requires not only the ability to stop them when problems occur, but also information to understand what is happening and to decide whether a stop is warranted. On November 3, CrowdStrike announced that it had signed definitive agreements to acquire Fiddler AI, which specializes in AI observability, and Dreadnode, which develops AI security testing. The Next Web also reported separately on the two planned acquisitions and on the AI Kill Switch, which is already available.
The features announced have different delivery timelines. Cutting off model connections via the Gateway is available now, while the endpoint-side shutdown function will come later. Meanwhile, "Falcon Observe," which uses Fiddler AI's technology, and automated AI security testing based on Dreadnode's technology are both scheduled to become generally available in the fourth quarter of CrowdStrike's fiscal year 2027 (FY27 Q4).
| Feature | What it controls or checks | Availability as of November 3, 2026 |
|---|---|---|
| AI Kill Switch (Gateway) | Cutting off AI model connections through the Gateway | Available |
| AI Kill Switch (endpoint) | Stopping agents running on endpoints | Planned |
| Falcon Observe (from Fiddler AI) | Tracking agent decisions and tool calls; pre-deployment evaluation | Scheduled for GA in FY27 Q4 |
| Automated AI security testing (from Dreadnode) | Testing models, applications, and more using known attack techniques | Scheduled for GA in FY27 Q4 |
This table organizes the target and availability of each feature and is not a performance comparison. FY27 refers to CrowdStrike's fiscal year and does not mean the end of calendar year 2027. Both acquisitions are at the signed-agreement stage, and their completion and the delivery timing of new features may change.
According to CrowdStrike's announcement on the Fiddler AI acquisition, Falcon Observe, based on Fiddler's technology, is envisioned to evaluate AI agents' accuracy, task completion, and policy compliance before deployment, and to track their decisions and tool calls once they are in operation. CrowdStrike also expects the acquisition to close in FY27 Q4, subject to customary closing conditions. Such tracking features would provide material for administrators to judge whether an agent should be stopped. However, even if the history of an agent's actions can be traced, it cannot necessarily be determined automatically whether those decisions were appropriate for the business.
Meanwhile, the announcement on the Dreadnode acquisition describes plans for automated testing using more than 100 known AI attack techniques against AI models, reasoning processes, applications, MCP servers, and more. CrowdStrike also outlined a vision of feeding the results of this testing into Falcon Guardian's policy recommendations. The aim is to verify AI systems' weaknesses before deployment and connect the findings to protections during operation, but this integration is a future plan, not a feature available today.
Companies considering the AI Kill Switch should first confirm whether their AI agents connect to models through the Gateway. They then need to determine whether cutting model connections is enough or whether halting execution on endpoints is also required. When behavior tracking and security testing, which inform the decision to stop an agent, will become available is also important. The announcement points toward managing AI agents in an integrated way, from monitoring and shutdown to pre-deployment verification, but it is essential to evaluate separately the features available now and those planned to be added after the acquisitions.