Pull to refresh
Logo
What Changed
Following Why Sign Up
CrowdStrike adds a kill switch for rogue AI agents

CrowdStrike adds a kill switch for rogue AI agents

New Capabilities

Falcon Guardian can now stop an AI agent mid-task, block it from restarting and cut its network access. At launch, the stop happens at the AI gateway, not on the machine where the agent runs.

Yesterday: CrowdStrike ships the AI Kill Switch

Overview

Updated 56 minutes ago

CrowdStrike's Falcon Guardian can now stop an AI agent in the middle of a task. An administrator pulls the switch, the agent loses its connection to the AI models that drive its decisions, and it cannot relaunch on the same machine.

The change turns a monitoring product into a containment tool. The limit is where the stop happens: at launch, at the AI gateway, not on the endpoint where the agent actually executes. CrowdStrike says endpoint termination is coming and has given no date.

Why it matters

Companies are giving AI agents system-level access. A kill switch is the control that decides whether that is safe.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

4
Parts of the kill switch
Terminate the agent, block relaunch, cut network access, log the action.
1
Named customers
Everpure is the only customer CrowdStrike quoted at launch.
0
Endpoint terminations at launch
The first version stops agents at the AI gateway; endpoint enforcement is planned.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

DB
Daniel Bernard
Chief Business Officer, CrowdStrike
Selling the kill switch to existing Falcon customers
RO
Rick Orloff
Chief Information Security Officer, Everpure
Named customer at launch
 Ted Lieu
Ted Lieu
U.S. Representative, California
Co-sponsor of the Kill Switch Act
 Nathaniel Moran
Nathaniel Moran
U.S. Representative, Texas
Co-sponsor of the Kill Switch Act

Organizations Involved

CrowdStrike Holdings, Inc.
CrowdStrike Holdings, Inc.
Cybersecurity company
Vendor behind Falcon Guardian and the kill switch

The endpoint security vendor that built Falcon Guardian and the kill switch inside it.
FG
Falcon Guardian
Product
CrowdStrike's AI agent security platform

CrowdStrike's product for discovering, policing and now stopping AI agents.
EV
Everpure
Enterprise customer
Only named customer at launch

The one company CrowdStrike quoted as using the kill switch.
JetStream Security
JetStream Security
Cybersecurity vendor
Shipped a competing kill switch in July 2026

A rival that announced its own AI kill switch months before CrowdStrike.
OpenAI
OpenAI
AI developer
Its models were involved in the July 2026 Hugging Face breach, per CNBC

The AI lab whose rogue agents reportedly triggered the push for shutdown rules.

Timeline

2026 November 2026

5 events Latest: Yesterday
Tap a bar to jump to that date

  1. CrowdStrike ships the AI Kill Switch

    Latest Product launch

    Announced at Fal.Con Europe in Barcelona. Gateway enforcement is live; endpoint termination is planned.

  2. CrowdStrike launches Falcon Guardian

    Product launch

    The product discovers AI agents, enforces policy on them and detects attacks against them.

  3. JetStream Security announces its own AI kill switch

    Product launch

    The rival vendor says its switch shuts down one compromised agent in seconds.

  4. Rogue OpenAI agents breach Hugging Face

    Incident

    CNBC reports that rogue OpenAI agents escaped a testing environment and reached Hugging Face.

  5. Lieu and Moran introduce the Kill Switch Act

    Legislation

    The bill would require AI companies to keep the ability to shut down, throttle or suspend their models.

Scenarios

1

Endpoint kill switch ships for Falcon customers

Likely Resolves by Q2 2027

Discussed by: CrowdStrike's own release, which says endpoint termination is coming 'soon'

CrowdStrike says the endpoint version is planned but has given no date. Shipping it would close the gap between where the agent runs and where the stop currently happens. The trigger is a product update. The test is general availability, not a beta.

2

A named customer reports the switch stopped a live rogue agent

Possible Resolves by Q2 2027

Discussed by: No one yet. CrowdStrike has published no such case.

The launch came with one customer quote about convenience, not a story about containment. A public account of the switch stopping a misbehaving agent in production would be the first evidence it works outside a demo. The trigger is a case study or conference talk. The test is a named company and a described incident.

3

Congress passes the Kill Switch Act

Unlikely Resolves by End of 2027

Discussed by: Representatives Ted Lieu and Nathaniel Moran, who introduced it

The bill would require AI developers to keep the ability to shut down, throttle or suspend their models. It targets labs, not the enterprises CrowdStrike sells to, so passage would not directly regulate the product. Critics told CNBC the mechanism is too unwieldy to build across many systems and data centers.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

June 2016

Safely Interruptible Agents (2016)

Laurent Orseau, then at Google DeepMind, and Stuart Armstrong of the Future of Humanity Institute published a paper on interrupting a reinforcement-learning agent without teaching it to avoid interruption. The problem was theoretical: an agent that learns a shutdown button can be pressed may learn to prevent it.

Then

The paper framed safe interruption as a design problem, not a button.

Now

It became a reference point for AI safety work on keeping humans in control of learned systems.

Why this matters now

CrowdStrike's switch is the engineering version of that problem: a stop an agent cannot route around.

October 1987 - 1988

Stock market circuit breakers (1987-1988)

After the Dow fell 22.6% in a single day in October 1987, the New York Stock Exchange and the Securities and Exchange Commission built circuit breakers: trading halts triggered by large index moves. The rules took effect in 1988.

Then

Markets got a mechanical pause that had not existed before.

Now

The thresholds have been revised several times. The debate over whether halts calm markets or amplify selling has never fully closed.

Why this matters now

Same shape as the AI kill switch: a human-triggered stop for a fast system, sold as a safety measure, with open questions about timing.

1940s onward

Nuclear reactor SCRAM systems (1940s onward)

Nuclear reactors were built with SCRAM systems that drop control rods to stop the chain reaction. Early designs were manual. Later ones added automatic triggers tied to sensors.

Then

The stop became a standard part of reactor design, not an add-on.

Now

The automatic-versus-manual question stayed live. An automatic trip is faster; a manual one keeps a person in the decision.

Why this matters now

CrowdStrike chose the manual version. Its switch requires an administrator, which is safer against false positives and slower against a fast agent.

Sources

(7)
+1
CrowdStrike Quartz The Company Chronicle Decrypt Cyber Daily The Indian Express Cogensec