Skip to main content

Enforcement

What is an AI kill switch?

An AI kill switch is an emergency control that stops an AI agent from continuing to act. For an agent, that means more than shutting down the model: the agent may still reach business systems through tool calls, scheduled jobs, and downstream automations. A working kill switch cuts off inference, the action surface, or both, and records what the agent did before the stop.

Why is a kill switch for an agent not one switch?

People say "turn off the agent" as if every deployment has one power button. An agent can keep affecting a business process through several paths: model inference, tool calls, scripts, scheduled jobs, and automations it triggered earlier. Stopping one of them does not necessarily stop the others.

Control pointWhat it stopsWhat to check
Inference The agent's ability to keep generating outputs and decisions. Can a separate process still execute queued or scheduled work?
Action surface The agent's ability to call governed tools and operations. Are all routes to business systems covered, including ones that bypass the gateway?
Both Inference and every governed action path. What operational work pauses, and who decides when to restore it?

No single setting is a universal off switch. The right response depends on the agent's architecture and the paths it can use, which is why those paths have to be mapped before an incident, not during one.

What is the difference between killing and quarantining an agent?

Killing an agent terminates it. That stops the behavior, but it can also interrupt a production workflow halfway through, and that interruption has its own cost: a half-processed batch, an order stuck between systems, a customer left waiting.

Quarantine freezes the agent's authority instead. The agent can no longer act, but its state is preserved, so people can investigate what happened and decide whether to restore, rescope, or retire it. In plain terms: stop the agent from acting, then work out the cause, without assuming that stopping inference closed every door.

How is LangGuard Agent Quarantine different from the usual ways to stop an agent?

Most teams improvise a kill switch from controls built for something else: rotating an API key, disabling an identity, blocking a route at the gateway, or killing the process. Each one stops part of the agent and misses the rest.

ApproachWhat it stopsWhat it misses
Revoke the model API key New inference calls. Queued and scheduled work, and tool credentials that are still valid. Takes down every other agent sharing the key.
Disable the identity New sign-ins and token issuance. Tokens already issued stay valid until they expire. A shared identity takes down everything that depends on it.
Block at the gateway Traffic routed through that gateway. Any path to a tool that does not go through it.
Kill the process The running agent. Its state, the workflow it was halfway through, jobs running elsewhere, and any record of why it was stopped.
LangGuard Agent Quarantine Inference, the action surface, or both, for one agent. Nothing on the governed action surface. State is kept for the investigation.

What makes the difference:

  • It knows what to stop. Quarantine works from the same action-surface map SCOPE-MCP keeps for every agent, so it targets that agent's tools and operations rather than a key or an account that other agents share.
  • It covers every path, not one choke point. Arbiter enforces at the AI infrastructure layer, the gateway and MCP catalogue, and inside the agent harness. A quarantined agent is refused at each of them, including routes that never touch the gateway.
  • It runs outside the agent. The agent does not have to cooperate, read a flag, or still be behaving correctly for the stop to hold.
  • It is granular. Stop inference, the action surface, or both, in real time, and restore it the same way once the cause is understood.
  • It fires from where incidents are handled. SOC teams can invoke it from a SIEM, a ticketing system, MCP, the API, or the LangGuard console, so it can be wired into existing playbooks rather than a separate tool someone has to remember.
  • It leaves a record. The audit ledger already holds every action and policy decision leading up to the quarantine, so the evidence for the review exists the moment the agent is stopped.
  • It works across platforms. The same control applies whether the agent runs on Bedrock, Anthropic, Databricks, or something built in-house.

Why do you need an inventory before you need a kill switch?

A kill switch is only as useful as your understanding of what it has to stop. The first question in any incident review is what the agent had access to and what else it could reach. If you cannot answer that, you cannot make a reliable decision about what to isolate.

An agent inventory is not a spreadsheet of project names. It ties each agent to its owner, its identity, the tools it is connected to, the specific operations those tools expose, the data behind them, and the downstream workflows it can trigger. That is the agent's action surface, and it changes every time someone connects a new tool. LangGuard's SCOPE-MCP maps it continuously, so the map is current when you need it.

How does managed identity reduce the need for a kill switch?

A kill switch is an emergency control. Least privilege is the everyday control that limits how often you need one.

  • Give each agent its own managed identity. Do not let it inherit a person's broad permissions because the workflow runs on that person's behalf.
  • Scope that identity to operations, not connections. Permission to connect to a system is not permission to perform every action in it. An accounts-payable agent may need to read invoices and prepare a payment for review. It does not need to change vendor bank details and release the payment.
  • Make the access revocable. Quarantine is fast and targeted when the identity is the agent's own. When the identity is shared or broad, pulling it takes down everything else that depends on it.

Most routine risk is handled by pre-execution enforcement, which blocks or escalates individual actions before they run. The kill switch is for when the agent itself, not a single action, is the problem.

What will a reviewer ask after the agent is stopped?

An incident review does not end with "we stopped it." A regulator, auditor, or customer will want to know what happened before the stop, who or what had authority, and what evidence supports your account. Be ready to answer:

  • Did we have an inventory of the agent and its connected tools?
  • Which operations and systems could it reach?
  • Which identity did it use, and what permissions did that identity hold?
  • Who could quarantine it, and through which response path?
  • What did it do before the control took effect?
  • Can we replay the action and authorization record?

A stop without an evidence trail leaves those questions open. LangGuard Arbiter evaluates actions before tool calls execute, and the audit ledger records each policy decision, so the record of what was allowed, what was blocked, and when the agent was stopped already exists. See provable authorization for what that record needs to contain.

How do you make an AI kill switch work before you need it?

  1. Inventory every agent and its action surface. Record the owner, tools, operations, systems, and downstream workflows.
  2. Assign a managed identity to each agent. Scope it to specific operations and remove inherited access it does not need.
  3. Connect quarantine to incident response. Decide how your SOC invokes it: SIEM, ticketing system, MCP, API, or console.
  4. Choose the response in advance. Set criteria for stopping inference, disabling the action surface, or both.
  5. Record the event. Capture which agent was stopped, who or what initiated it, when, what actions preceded it, and which policy applied.

Common questions

Is shutting down the model the same as killing the agent?

No. Stopping inference stops the agent from generating new decisions, but work it has already queued, scheduled jobs it created, and automations it triggered can keep running. An agent that calls tools through more than one path may also have routes that never pass through the model provider you shut off. Stopping an agent reliably means cutting off its action surface as well as its inference, and checking which processes were still running when you did.

Why not just revoke the agent's credentials?

Revoking credentials stops some paths and leaves others open. Disabling an identity blocks new tokens, but tokens already issued can stay valid until they expire. Rotating a model API key stops inference but not work that is already queued, and it takes down every other agent using the same key. LangGuard Agent Quarantine acts on one agent across its whole governed action surface, enforced at the gateway, the MCP catalogue, and the agent harness, and it preserves the agent's state and decision record for the investigation.

Who should be able to trigger an AI kill switch?

Whoever handles the incident, through the tools they already use. In most organizations that is the security operations team working from a SIEM or ticketing system, not the engineering team that built the agent. The control should be reachable from those systems, with the agent's owner and the business process owner told when it fires, because they decide when the agent can be safely restored.

Does a kill switch replace pre-execution enforcement?

No. They work at different levels. Pre-execution enforcement evaluates each action before it runs and blocks or escalates the ones that fall outside policy, which handles routine risk without stopping the agent. A kill switch acts on the whole agent and is for when the agent itself is the problem, for example after a compromise or when its behavior no longer matches what it was approved to do. A deployment with good per-action enforcement needs the kill switch less often.

Are AI kill switches required by regulation?

For high-risk AI systems, the EU AI Act's human oversight requirement in Article 14 includes the ability to interrupt the system through a stop button or similar procedure that brings it to a safe halt. Elsewhere the direction is similar but less settled. In November 2026 the Australian government accepted recommendations that include considering mandatory AI incident reporting and announced plans for a new AI safety regulator; those are proposals, not enacted law. In both cases the expectation goes beyond stopping the system: organizations are increasingly expected to explain an incident and show the controls they had in place.

Know what your agents can reach before you need to stop them.

LangGuard maps your complete agent action surface in minutes. Free for the first five managed agents, with every Scope included from day one.