CrowdStrike launches AI kill switch as it pursues a reported $400 million Fiddler deal
The new feature is designed to stop misbehaving internal AI agents and restrict their access. CrowdStrike is also pursuing two AI-security acquisitions, with reported prices for both deals and a proposed fiscal 2027 closing.

The Wall Street Journal reported this proposed price for Fiddler AI; CrowdStrike has not publicly disclosed the price.24
The Wall Street Journal reported this proposed price for Dreadnode.4
CrowdStrike expects both proposed acquisitions to close then, subject to customary closing conditions.4
Story
A human-controlled stop for enterprise AI agentsCrowdStrike introduced an AI Kill Switch at its Fal.Con Europe conference on Tuesday, adding a way to halt internal AI agents that behave improperly.1 The company says the feature can prevent an agent from starting again and remove its access to an enterprise network.1 The launch is part of Falcon Guardian, a platform CrowdStrike released in September to protect AI agents.1 At first, the shutdown operates through AI gateways, which mediate an agent’s access to the models that inform its decisions.2 CrowdStrike has described secure operation of AI agents as a substantial business opportunity, making the release both a security measure and part of its wider product effort.1 Because the current action works through gateways, the launch provides a way to interrupt access to underlying models, rather than the full range of containment CrowdStrike says it plans to add later.2
The initial release does not directly terminate an agent process running on an endpoint, nor does it coordinate restrictions across cloud services and other enterprise systems.2 CrowdStrike lists those capabilities as later extensions rather than features already available.2 This distinction defines what the present control can do: it can cut off model access through a gateway, while several broader forms of intervention remain planned.2 When the control is used, a security administrator starts the action, and the event is recorded in an audit trail.2 The recorded action provides a trace of the intervention, while the decision to initiate it rests with an administrator.2 For organizations relying on agents, the gap between gateway-based intervention and direct endpoint or cross-system restrictions marks the limits of the initial release.2
CrowdStrike says Falcon Guardian can examine agent prompts and apply rules to several coding and workplace assistants.2 The named products include Claude Code, OpenAI Codex, GitHub Copilot, Microsoft 365 Copilot, Cursor and OpenCode.2 The listed scope shows that the product is intended to manage activity involving multiple agent tools, rather than a single named assistant.2 Daniel Bernard said the system identifies activity by internal agents, but does not itself decide whether that activity is harmless or malicious.1 That division leaves the judgment about an alert with a person, consistent with the administrator’s role in starting the recorded shutdown action.12 The combination of prompt inspection and human judgment means the system can provide oversight across the named tools without making the final benign-or-malicious determination itself.12
CrowdStrike CEO George Kurtz said the feature can stop an agent from restarting and sever its connection to the company network.1 He also presented the control as a way to give people greater authority over AI systems as worries about managing agents increase.1 Bernard described the ability to govern agents and switch them off as important to the next stage of AI adoption.1 Kurtz likened runtime protections to airbags in cars and said CrowdStrike could stop agents that go beyond their boundaries.1 Together, these comments frame the kill switch as an intervention for use while agents are operating, not just a way to assess them before use.1 The emphasis on stopping a running or restarting agent reflects CrowdStrike’s argument that control mechanisms are important as AI adoption advances.1
CrowdStrike also announced that its products would integrate with Anthropic’s Claude Cowork.1 Kurtz said enterprise customers had asked Anthropic to include CrowdStrike technology before they put Anthropic products into use.1 That account links the integration to requests from customers considering Anthropic products, while the announcement itself identifies Claude Cowork as the integration.1 CrowdStrike’s other recent effort in this area was a Falcon control plane for AI agents, announced in June.4 The company said that control plane continuously applies security decisions according to an agent’s identity and current risk.4 The Claude Cowork integration and the earlier control plane are separate parts of CrowdStrike’s efforts to bring security controls to AI-agent use.14
CrowdStrike has signed agreements to buy Fiddler AI, which works on AI observability, and Dreadnode, a company focused on testing AI security.1 The Wall Street Journal reported proposed prices of about $400 million for Fiddler and about $95 million for Dreadnode.4 CrowdStrike has not announced Fiddler’s price publicly; Decrypt reported that the consideration is largely cash, with some stock subject to vesting terms.2 The reported amounts are descriptions of proposed deal values, not prices CrowdStrike has confirmed publicly for Fiddler.24 Both purchases remain proposed transactions, and CrowdStrike expects them to close in the fourth quarter of fiscal 2027 if customary closing requirements are met.4 The reported Fiddler amount is substantially larger than the reported Dreadnode amount, though both deals remain subject to closing conditions.4
Fiddler’s software can assess agents before they are deployed, follow their decisions and tool use, and track changes in their behavior afterward.2 Dreadnode offers automated security testing that lets companies check AI systems for vulnerabilities before deployment.4 CrowdStrike says it plans to present Fiddler’s capabilities as Falcon Observe.3 It also plans to make the capabilities of both companies available across Falcon by early 2027.3 The acquisitions would extend Falcon’s controls beyond the identity-and-risk authorization capability announced in June.14 Fiddler’s tools cover agent assessment before deployment and behavioral tracking afterward, while Dreadnode provides pre-deployment vulnerability testing; Falcon’s control plane applies identity and risk decisions continuously.24
CrowdStrike says enterprise agents can carry out tasks involving several steps across company systems without direct human oversight.3 The company says that capacity increases the security demands associated with using such agents.3 President Mike Sentonas identified poor configuration, gaps in user understanding and overly broad permissions as important sources of risk.3 Those concerns sit alongside the product’s human decision point: the system can detect internal agent activity, while a person judges whether it is malicious or benign.13 The launch therefore combines monitoring and policy enforcement with an administrator-controlled action, while the wider endpoint and cross-system controls remain planned.12 The risk factors Sentonas named concern how agents are configured, understood and authorized, while the kill switch addresses the ability to intervene in their activity.13
The debate over controls extends beyond CrowdStrike’s product announcements.1 In September, Anthropic CEO Dario Amodei argued for slowing AI development because of safety concerns, and OpenAI CEO Sam Altman supported that position.1 The CNBC Technology article also reported that U.S. representatives introduced bipartisan legislation over the summer that would require AI companies to include shutdown functions in their models.1 Against that backdrop, CrowdStrike’s release is a company-level control that currently works through AI gateways, while the reported legislation concerns shutdown capabilities built into models.12 CrowdStrike’s stated approach places the immediate decision with a human administrator and treats broader containment across endpoints and enterprise environments as future work.12 The comparison is between a gateway-based product feature available now and a proposed requirement for shutdown functions in AI models; the two address control at different levels.12
Structure
Who is connected to whom- 1CrowdStrike
- partners with →Fact
- produces →Fact
- ← suppliesFact
- geographic exposure →Fact
- uses →Fact
History
How it came to this- JuneFalcon control plane announcedCrowdStrike announced a control plane for AI agents that continuously enforces actions based on identity and real-time risk.
- SeptemberFalcon Guardian launchedCrowdStrike launched the platform to secure AI agents; the AI Kill Switch is built into it.
- TuesdayAI Kill Switch launchedCrowdStrike introduced the feature at its Fal.Con Europe conference.
- 2026-11-04Earlier acquisition signalsEarlier signals flagged CrowdStrike's two AI-security startup acquisitions and plans to enhance AI security.
- NowTuesday CrowdStrike launches AI Kill Switch
- Early 2027CrowdStrike said it plans to make both companies' capabilities available across Falcon by early 2027 and to offer Fiddler capabilities as Falcon Observe.
Impact
Spreading outward, level by level- Level 1Immediate agent intervention
A security administrator initiates the action, which is recorded in an audit trail. In the initial release, the kill switch is available through AI gateways that block an agent's access to the models behind its decisions.2
Fact - Level 2Limits of the initial release
Direct termination of agent processes on endpoints and coordinated restrictions across cloud and other enterprise systems are planned expansions, not currently part of the initial release.2
Fact - Level 3Human oversight
The system detects internal agent activity, but a person decides whether it is benign or malicious; the administrator also initiates the shutdown action.12
Fact - Level 4Broader Falcon controls
The proposed acquisitions could extend Falcon's controls beyond the identity-and-risk authorization capability announced in June, addressing agent security before deployment and during operation.
Analysis
Ahead
Checked automatically when due; the result goes to the track recordCrowdStrike said it plans to make both companies' capabilities available across Falcon by early 2027 and to offer Fiddler capabilities as Falcon Observe.
CrowdStrike expects the proposed Fiddler AI and Dreadnode acquisitions to close in the fourth quarter of fiscal 2027, subject to customary closing conditions.
Sources
What each source supportsWritten by AI from the sources listed below: every fact was checked word for word against its source, and inference is marked apart. How we write