CrowdStrike buys Fiddler AI and Dreadnode for $495 million in connection with launch of AI Kill Switch
CrowdStrike launched an "AI Kill Switch" at Fal.Con Europe in Barcelona on November 3, 2026 — and at the same time signed agreements to acquire two AI security companies, Fiddler AI and Dreadnode, for a combined $495 million according to the Wall Street Journal. The feature is intended to give security teams a way to stop AI agents behaving unexpectedly. But what the feature actually does today is far more limited than the name suggests: For now, it works only through connected AI gateways that cut the agent's access to the models it uses. Coverage for devices, cloud-based agents, permissions, and browser operations is still "coming soon." Within the same week, Microsoft (November 5) and OpenAI (November 9) rolled out their own control mechanisms for AI agents — a concentrated vendor offensive showing that the "off button" for autonomous agents has become a competitive battleground, not a curiosity.
What Was Actually Launched
The announcement came in a press release on November 3 in connection with the Fal.Con Europe conference in Barcelona. The press release describes the AI Kill Switch as a feature that gives security teams "the ultimate human control over AI agents," and states that "The AI Kill Switch is available today" — that is, available from day one, not just a vision (24/7 Wall St.).
The AI Kill Switch launched as one of several new agent-related capabilities: CrowdStrike simultaneously introduced Agentic Adversary Emulation powered by SafeMind as well as Agent Insider Threat Detection, and announced ecosystem integrations with, among others, AWS, Anthropic, Dynatrace, and AppGate (Simply Wall St. via Yahoo Finance). The picture, then, is not a single feature but a broader packaged offering aimed at security around AI agents.
The same day, the company signed agreements to acquire two AI security companies: Fiddler AI, which develops tools that help companies track and monitor their AI systems, and Dreadnode, which develops tools for testing AI systems for security vulnerabilities. The combined purchase price is $495 million, according to the Wall Street Journal (Yahoo Finance, 24/7 Wall St.). There are no details on deal terms, timeline, or any regulatory review in the source material, and such questions should be left open until CrowdStrike or the parties themselves provide more information.
How the Mechanism Works — and What It Doesn't Do Yet
The concrete mechanism today can be described as what CrowdStrike itself calls an "agent circuit breaker." When security teams activate the feature, it is connected AI gateways that block the agent from reaching the models it needs to make decisions. In other words: the agent no longer gets responses from the language model, and thus cannot continue acting on its own (MSN).
That is an important nuance: the blocking happens at model access, not by killing the agent process itself on the machine. This means the feature today presupposes that agent traffic flows through a gateway CrowdStrike has visibility into. An agent not connected to such a gateway cannot be stopped by this mechanism as it currently works.
CrowdStrike itself has listed what is available and what is coming (note.com, summarizing the announcement):
- AI gateway (cuts the connection to AI models): available now
- On-device (stops the agent process and prevents restart): coming later
- Cloud-based agents: coming later
- Agent permissions (revoked immediately): coming later
- Enterprise cloud services and browser operations: coming later
The company plans to extend the feature to also cover endpoints, cloud environments, and enterprise identity systems, in order to stop "rogue" AI agents more broadly (MSN). Yahoo Finance's retelling of the announcement further suggests that a defender will be able to terminate the agent "where it runs," prevent it from being restarted, and cut its network access (Yahoo Finance) — but the more specific sources are unambiguous that this today applies only through connected gateways, and that endpoint and cloud coverage are future extensions. That is why the accurate answer to "what can we use today?" is gateway-based blocking of model access, not a universal off switch across devices.
What Kurtz Says — and How It Should Be Read
CrowdStrike founder and CEO George Kurtz has framed the launch as a matter of human control: "AI safety is how you get the full benefit of AI. When you deploy agents with a kill switch, you stay in control," he said in the press release (Yahoo Finance). On CNBC's Mad Money, he reportedly said: "Everyone is building AI agents, but they never had the off button — now they do" (note.com, as reported via CNBC).
Both quotes are company statements conveyed through secondhand sources, not independent assessments. It is worth keeping them separate from what is technically documented: That agent control is a real and growing need is a reasonable position in light of multiple major vendors and government agencies moving into the same field. That CrowdStrike's specific implementation is the best or most complete is not something the sources can confirm.
The Competitive Picture: Three Vendors in One Week
The launch did not happen in a vacuum. Within days, two directly competing offerings followed:
Microsoft (November 5) detailed Dynamic Permissions — which grants an agent authority only for the task at hand — along with Dynamic Agent Risk and Conditional Access, which can tighten permissions or require human approval (MSN). Microsoft's approach is fundamentally different from CrowdStrike's: it works with permissions before and during operation — what rights the agent gets, and when it must stop and ask for approval — rather than cutting access to models afterward. It is a preventive rather than reactive move.
OpenAI (November 9) rolled out its own controls for agents, according to the same source material. The details on these are thinner in the available sources, so what can be said with confidence is that OpenAI is the third major player to publicly position itself on agent control in the same week (MSN).
The backdrop is clear: every vendor wants to become the control point for how organizations govern AI agents. CrowdStrike's model logically fits into an existing security team already using the Falcon platform, while Microsoft can leverage its position in the enterprise and its identity/permission system, and OpenAI can build control in where the agents actually run. For buyers of these products, that means the choice of kill switch could become a vendor platform battle — and interoperability between them remains an open question.
The Reference Point: The October NIST–NSA–CISA Guide
The government level is also represented: In October, NIST, NSA, and CISA published a guide on practices around "rogue" AI agents. The guide sets voluntary recommendations, not requirements — but it functions as a common reference point for what vendor features like kill switches and dynamic permissions are actually meant to solve (MSN). Three vendors rolling out mechanisms in the same week, right after such a guide, suggests the industry is at least synchronizing around the same problem — regardless of whether the features are technically equivalent.
Why the Acquisitions Fit: Observation and Testing Around the Switch
The two acquisitions give a picture of what CrowdStrike is building around the kill switch itself. Fiddler AI develops tools for observability and tracking of AI systems — that is, what makes it possible to detect that an agent is behaving abnormally before anyone needs to press the switch. Dreadnode develops tools for testing AI systems for security vulnerabilities (Yahoo Finance).
This is not coincidental: a kill switch is only useful if someone knows when to activate it. Without observation and testing around it, the switch becomes a backup tool you hope you never need. With Fiddler and Dreadnode, CrowdStrike is building the whole chain — detect, test, and stop — and it is presumably the broader platform position, not the switch alone, that is the strategy.
The context makes the investment look far from marginal: CrowdStrike reported $5.84 billion in annual recurring revenue (ARR) at the end of the second quarter of its fiscal year 2027 (ended July 31), up 25% from the same period a year earlier. The company's ARR for its AI security products (AIDR) reportedly nearly tripled quarter over quarter — all figures company-reported (Motley Fool). That means AI security is already a growing revenue source, not a future scenario, and that CrowdStrike is putting capital behind a category where it is itself seeing demand.
What This Means for Organizations Considering Agent Control
For organizations already using CrowdStrike Falcon and running AI agents through gateways the platform can see, the AI Kill Switch can be a concrete, usable tool today — limited to cutting model access, but that covers many practical situations where an agent "runs loose" because it keeps pulling decisions from a model.
For organizations with agents running on-device, in the cloud, or via the browser without a connected gateway, the feature today offers no guarantee. CrowdStrike itself states that this coverage is coming later. That means a realistic assessment today should be based on gateway coverage, and treat the broader containment offering as a planning assumption, not an existing capability.
The comparison with Microsoft is also relevant: if an organization finds that dynamic permissions and risk-based controls — that is, preventing an agent from getting authority it doesn't need — are better suited than cutting model access after the fact, it is not given that CrowdStrike's model is the best approach. Presumably, a coherent strategy would combine both: restrict permissions up front, and have a switch as a last resort. But which products can interoperate across vendors is not something the sources can confirm.
Open Questions
Several things are not answered by the available sources, and that affects how confidently the reporting can be trusted so far:
- Deal terms, timeline, and regulatory review for the Fiddler AI and Dreadnode acquisitions are not documented anywhere in the sources, and should not be assumed until CrowdStrike itself provides updates.
- The actual product documentation from CrowdStrike is not available here; all capability descriptions come via media outlets reproducing the press release or curated summaries. The technical details — such as exactly which gateways are supported, the latency of blocking, and how the agent is perceived after a block — are not independently verified.
- The discrepancies between sources' descriptions are clear: one source suggests a defender can stop the agent "where it runs," while the more specific sources limit the feature to gateway blocking today. The more specific and limited description should be given the greatest weight until anything else is documented.
- The market reaction is reported inconsistently across sources — from shares "up more than 1%" to smaller moves after market — and should be treated with caution rather than as a fact.
- Interoperability between vendors is not covered: it is unclear whether an organization using CrowdStrike, Microsoft, and OpenAI agents alike can tie the controls together, or whether each vendor expects to be the exclusive control point.
What remains after one hectic week is a clearer market picture than a technical scorecard: CrowdStrike is positioning itself early with a concrete but still limited off switch for AI agents; Microsoft is betting on permissions and risk management; OpenAI is entering the field from the vendor side of the agents themselves. For potential users, the most important thing right now is to read the launches specifically — what the feature actually does today, not what it will do in a future release.

