How do you stop AI agents across an organization?
By Verillian. Published . Updated .
Limit what an agent may do with managed settings and policy, and keep one kill switch for every enrolled device. What Anthropic’s docs say for Claude Code.

It takes two controls that do different jobs. One limits what an agent may do, ahead of time, on every machine. The other denies AI requests on every device when something has gone wrong. For Claude Code, Anthropic ships the first in managed settings. The second has to sit outside any one tool.
Can Claude Code’s settings be enforced across an organization?
Yes. Anthropic’s documentation describes managed settings as “settings your organization deploys, by a managed-settings.json file, an MDM policy, or server-managed settings from the claude.ai console,” and says “Nothing in your own settings files or --settings overrides a managed key.”
The rules inside them are firm. In the permissions documentation, “Rules are evaluated in order: deny, then ask, then allow,” and “An allow rule can’t carve an exception out of a deny rule.”
Bypass mode can be turned off with permissions.disableBypassPermissionsMode, and Anthropic says this is most useful “in managed settings where they can’t be overridden.” The check doesn’t depend on the model’s judgment either. “Permission rules are enforced by Claude Code, not by the model.”
What don’t managed settings do?
Anthropic is plain about the edges. A rule on a command “covers the invocation Claude usually produces and isn’t a security boundary around the program.”
Read and edit rules don’t reach “arbitrary subprocesses that read or write files indirectly, like a Python or Node script that opens files itself.” And “a managed-settings.json file or MDM profile on your device doesn’t reach a cloud session.”
Settings also belong to one tool. Every other agent on the same device has its own, or none, and a rule written in advance isn’t a way to stop them all when something goes wrong.
Our post on destructive commands sets out the layers that close the gap on commands, from least privilege and approval to a check at the moment of the call.
What should a kill switch do?
A kill switch has to reach every device whichever tool is running, be set by a person, and leave a record of who set it and why. You’ll want to know what ran just before it, too.
State bank examiners now ask about it. The CSBS Core Examiner Guide reviews “the actions the system is permitted to take, human checkpoints, logging, reversibility, and the ability to restrict or halt the system.”
Lawmakers are asking too. A New York City bill, as drafted, would require a shut-down capability that a human operator can use.
How does Verillian contain AI agents?
In Claude Code, a tool call your policy denies never runs. The kill switch denies AI requests from every enrolled device, and a containment covers one person, device or provider. Each is recorded with who set it and when. The records of what ran before the kill switch stay on your servers.
Sources
- Anthropic, Claude Code settings documentation. Anthropic documentation.
- Anthropic, Claude Code permissions documentation. Anthropic documentation.
- Conference of State Bank Supervisors, Core Examiner Guide. CSBS Core Examiner Guide.
Sources read on 3 November 2026.
See Verillian govern AI on a real device.


