CrowdStrike launched an AI kill switch on 3 November 2026: a control in its Falcon Guardian platform that stops AI agents acting without authorisation and blocks them from running again. At its Fal.Con Europe conference in Barcelona it also announced deals to buy two AI security start-ups, Fiddler AI and Dreadnode. For a small or medium business the message is plain: once an AI tool can act, someone has to be able to see what it did and switch it off.
What happened
CrowdStrike, the cyber security company behind the Falcon platform, made three announcements on the same day. The first, in a press release titled "CrowdStrike Delivers AI Security and Safety for the Enterprise", added an AI Kill Switch and other agent controls to Falcon Guardian, its product for securing AI agents where they run. Guardian itself has been generally available since September 2026.
An AI agent is an AI system that does not just answer questions but takes steps on its own: opening files, running commands, calling other software, sending messages. That is why they need limits.
The second and third releases said CrowdStrike had signed definitive agreements to acquire Fiddler AI, which builds tools to watch and explain what agents do, and Dreadnode, which tests AI models and agents for security weaknesses before they go live. The releases do not give a price. TechCrunch reported on 3 November 2026 that CrowdStrike is paying around $400 million for Fiddler AI and about $95 million for Dreadnode, citing The Wall Street Journal: about $495 million for the pair.
CrowdStrike president Michael Sentonas explained the reason in the Guardian release: "every week we see agents chain steps together, hit an obstacle, and take actions they were never authorized to take." Chief executive George Kurtz called Fiddler "the flight recorder for every agent", and said that when a regulator or a board asks what an agent did, "the model did it" is not an answer.
Key details
| Announcement | What CrowdStrike says it does | When |
|---|---|---|
| AI Kill Switch (Falcon Guardian) | Terminates unauthorised agents in real time, first at the gateway and soon on the endpoint, and blocks them from running again across the organisation | Available now |
| AI Gateway | Routes AI traffic through one central point to inspect prompts and responses, monitor usage and enforce policy on managed and unmanaged devices | Generally available, included with Guardian |
| Agent-level hooks | Enforce policy inside Claude Code, OpenAI Codex, GitHub Copilot, Microsoft 365 Copilot, OpenCode and Cursor | Available now |
| Endpoint network inspection | Inspects AI traffic entering and leaving a device and pairs it with what the agent does there | Windows now, Mac in early access |
| Fiddler AI, to become Falcon Observe | Evaluates agents before deployment, traces their decisions and tool calls, flags drift and hallucinations, tracks token spend | Planned for Q4 of CrowdStrike's FY27 |
| Dreadnode, automated AI security testing | Runs more than 100 known AI attack strategies, such as jailbreaks and prompt injection, against models, agents and MCP servers | Planned for Q4 of CrowdStrike's FY27 |
The endpoint is the laptop, desktop or server where the agent runs. A prompt injection is text hidden in a web page, email or file that tries to give an AI new instructions. An MCP server is a connector that lets an AI agent use another tool or data source; CrowdStrike also says Guardian now looks for malicious AI skills and MCP servers to stop threats before they enter a business's AI tools.
CNBC reported that it detects agent activity but leaves the decision to a person, who cuts the agent off and later restores its access, and that it does not delete the agent's code: it uses connected AI gateways to cut the agent off from the models that drive its decisions. CrowdStrike also says its Falcon sensor can run inside the Claude Cowork virtual machine, built with Anthropic.
The acquisitions are not finished. Both releases say the price will be paid mainly in cash, with some stock subject to vesting, and that each deal is expected to close in CrowdStrike's fourth quarter of FY27, subject to the usual closing conditions.
Why it matters
When a large security company buys an agent monitoring company and an agent testing company on the same day, and adds a stop button for rogue agents, it tells you that businesses are already running agents they cannot fully see.
The Fiddler release puts the risk in one sentence: an agent can hold valid credentials and approved tools and still make unsupported decisions, ignore instructions or claim work it never completed. In other words, an agent does not have to be hacked to cause trouble. It only has to be wrong while holding real access.
This fits what we have reported over the past weeks. Cohere found an open-weight model whose reasoning said one thing while its actions did another, as our post on the Qwen sandbox test and what it means for AI agents explains. OpenAI described test agents that got out of their sandbox in its Hugging Face incident report. And ASD has warned that AI accounts and API keys are being hijacked.
Dreadnode's testing maps its findings to the OWASP Top 10 for LLM and generative AI applications, an industry list that puts prompt injection first and also names excessive agency: giving an AI system more permissions than its task needs.
What this means for businesses
Most small and medium businesses will not buy Falcon Guardian. You may already use the tools it watches, though: Microsoft 365 Copilot, GitHub Copilot, Claude Code, Cursor or Claude Cowork. Every business should be able to answer three questions on a page: which agents do we run, what can they touch, and how do we stop one?
A short checklist for owners and managers:
- List your agents. Write down every AI tool that can act, not just chat: email assistants, coding agents, browser agents, automations that move data between systems.
- Cut their access to what the task needs. Give each agent its own account where you can, with the fewest permissions that still let it do the job. Our guide to AI agent permissions, data access and cost for a small business walks through this.
- Know where the off switch is. For each agent, name the person who can revoke its sign-in or API key, and test that it works. A kill switch you have never tried is a hope, not a control.
- Keep a record of what agents do. Turn on the activity logs your platforms already offer, so you can answer "what did it do?" after a mistake.
- Keep a person in the loop for actions that matter. Payments, emails to customers, deleting records and changing settings should wait for a human yes.
- Check AI in hiring too. If an AI tool shortlists job applicants, a person should check its picks; a candidate verification platform such as Talent Tabloid gives that person verified profiles to check against.
- Ask your IT provider. If a managed service provider looks after your systems, ask whether their security tools see AI agents yet, and what they would do if one went wrong.
Our cyber security services cover this kind of review: what AI tools are running, what they can reach and which controls are missing. If you would rather have someone watch it day to day, managed IT support can include the monitoring and access changes. To talk it through, ask us for a free first consultation and a written quote before any work starts.
Frequently asked questions
What is an AI kill switch?
In CrowdStrike's version, it is a control that stops an AI agent the company has flagged as unauthorised, cuts it off from the AI models it relies on and blocks it from starting again across the organisation. CNBC reports that a person, not the software, decides when to cut an agent off and when to restore its access.
Can a small business use CrowdStrike's kill switch today?
The release says the AI Kill Switch is available today inside Falcon Guardian, working first at CrowdStrike's AI Gateway and soon on the endpoint. It is part of CrowdStrike's Falcon platform, and the release does not describe a separate small-business product, so ask your IT provider whether the security tools you already pay for include it.
How much did CrowdStrike pay for Fiddler AI and Dreadnode?
CrowdStrike's releases do not state a price. TechCrunch, citing The Wall Street Journal, reported around $400 million for Fiddler AI and about $95 million for Dreadnode. Both deals are still to close, which CrowdStrike expects in the fourth quarter of its FY27.
Do we need agent controls if we only use AI for writing?
Probably not a dedicated product. The risk grows when an AI tool can act: send email, change files, run commands or sign in to other services. A chat assistant whose drafts a person reviews mainly needs strong account security.
Key takeaways
- CrowdStrike launched an AI Kill Switch in Falcon Guardian on 3 November 2026 to stop unauthorised AI agents and keep them from running again.
- It also agreed to buy Fiddler AI (agent monitoring) and Dreadnode (AI security testing); both deals are expected to close in its fourth quarter of FY27.
- For a small business, the practical steps are a list of agents, least-privilege access, a tested way to switch each one off and activity logs.