One day, an employee at your organization notices that a co-worker’s Claude Code session is acting outside its intended scope. Or an agent approved for a pilot has been retired.
You need that single connection stopped, but you do not want to lock the employee out of every other agent, and you can’t risk breaking Claude for the rest of the team. At that moment, you need a way to shut off access for that agent alone without affecting the user or other agents using the same resources. You need, in common parlance, a kill switch.
An AI agent kill switch is an administrative control that revokes a user’s access through a specific AI agent, so security teams can contain a problem without disrupting other access. Aembit’s version is now available for customers using Aembit IAM for Agentic AI to secure their agents.
Why Is Containing One AI Agent So Hard?
Security teams usually have three options for cutting off an agent, and each has a limitation:
- Change an access policy. The change applies to every user of that MCP client, not just the one you are worried about.
- Deprovision the user at the identity provider. This removes the user’s access well beyond a single agent.
- Wait for tokens to expire. The agent stays authorized until the tokens expire.
Containing an agent problem takes a control tailored to the problem: one user, one agent.
This matters more as agents multiply. Every employee can now connect several AI agents to enterprise systems through MCP, and each pairing of user and agent is its own path to sensitive data. Visibility into those paths is necessary, but visibility alone does not stop anything. Teams also need a way to take action quickly and precisely when something looks wrong.
How Does Aembit’s AI Agent Kill Switch Work?
Aembit’s kill switch is a session revocation control that ends one user’s session through one MCP client. When a user signs in to an MCP client such as Claude Desktop, their identity provider authenticates them, and Aembit issues access and refresh tokens for that sign-in.
Aembit records each sign-in as an MCP user session and lists it in your tenant with the user’s email, the MCP client, when the session was issued, when it expires and whether it is active or revoked. An administrator selects a session and revokes it. That sign-in’s tokens stop working, so the agent can no longer use that session to reach MCP servers. The same user’s other agents keep working, and every other user keeps working in the same client.
To revoke a session:
- In your Aembit tenant, go to Administration, then Identity Providers, and open the provider your users sign in through.
- Open the Sessions tab and find the session by user email or MCP client.
- Select Revoke session and confirm.
Revoking a session requires Read/Write permission on Identity Providers. Sessions appear for MCP clients that connect through a Client-to-Gateway Access Policy that uses an OIDC ID Token Credential Provider with refresh token support enabled.
What Can You Revoke With the Kill Switch?
The same Sessions view lets you choose which sessions to revoke:
| What you want to stop | How to do it |
|---|---|
| One sign-in | Select the Revoke session icon on that row |
| One user’s sessions in one MCP client | Filter by user email and MCP client, then select Revoke filtered |
| All sessions for one user | Filter by user email, then select Revoke filtered |
| All sessions for an MCP client | Filter by MCP client, then select Revoke filtered |
Teams that automate incident response can also revoke sessions through the Aembit Cloud API.
When Should You Use an AI Agent Kill Switch?
A kill switch can help when:
- You suspect a user’s credentials or device have been compromised and want to cut that agent connection first.
- An agent behaves outside policy, and you need to stop its access while you investigate.
- A user’s approval for one AI tool has ended, but their other access should continue.
- An MCP client is retired or fails a review, and you want to revoke every session it holds.
How Does an AI Agent Kill Switch Fit With Access Policies and Your Identity Provider?
Your identity provider still authenticates people, and Aembit access policies still decide which MCP clients and servers they can reach. The kill switch adds a targeted, on-demand way to end a specific sign-in. After a session is revoked, the user can sign in again from the same client, creating a new session that your identity provider and policies evaluate as usual.
For lasting changes to who can use which client or server, update the access policy. The kill switch does not revoke a user’s downstream connections to individual MCP servers, and it does not deprovision anyone at the identity provider.
Is Aembit’s AI Agent Kill Switch Available Now?
The kill switch is available now for customers securing their AI agents with Aembit IAM for Agentic AI. Talk to an engineer to see it in your own environment, watch the demo or read the docs.
