Kill Switches and Managed Identities: What Every Agent Deployment Needs Before the Next Inquiry

Illustration of an AI agent, its connected action surface, and an isolation boundary

Kill switches are having a moment. Australia announced plans for mandatory AI incident reporting and a new safety regulator after an AI agent accessed a government Medicare statistics portal. In Europe, another round of information requests is bringing recent AI incidents and safety practices into focus.

The control conversation is right. But a kill switch is only a promise unless you know which agents you have, what each one can do, and how to cut off that authority when you need to.

A map, a least-privilege identity, and a quarantine control are not three alternatives. They are one control system.

You cannot stop what you cannot see

The Australian incident is a useful reminder of what an inquiry looks like. In June, an OpenAI agent accessed non-public Medicare statistics through a Services Australia portal. OpenAI reportedly detected the access in August and notified the government in September. On November 10, the Australian government accepted all 18 recommendations from a rapid review, including consideration of mandatory, time-bound reporting for AI incidents and misaligned behavior. It also announced plans for a new AI safety regulator and mandatory safety and security protections. The ABC has reported that no personal information was accessed.

That sequence raises a basic operational question: what did the agent have access to, and what else could it reach?

An agent inventory is not a spreadsheet of project names. It should connect each agent to its owner, identity, tools, operations, data, and downstream workflows. If you cannot see the action surface, you cannot make a reliable decision about what to isolate.

LangGuard’s Live AI Agents and Tools Registry and SCOPE-MCP map agents to the tools and operations they can reach. The job is practical: discover the connections, classify the available actions, and keep the map current as the environment changes.

Inventory first. A quarantine control is only as useful as your understanding of what it needs to quarantine.

Diagram of an AI agent mapped to its tools, data systems, APIs, and workflows

A kill switch is not one switch

People say “turn off the agent” as if every deployment has one power button. It may have several ways to keep affecting a business process: inference, tool calls, scripts, scheduled jobs, or downstream automations. Stopping one component does not necessarily stop the others.

There are two different control points:

Control point What it stops What to check
Inference The agent’s ability to continue generating outputs 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?
Both Inference and governed action paths What operational work pauses, and who decides when to restore it?

LangGuard’s Agent Quarantine adds a real-time option to disable an agent’s inferencing, its action surface, or both. Teams can invoke it through a SIEM, ticketing system, MCP, API, or the LangGuard GUI. The point is to make the control available where an incident is handled, not only in a separate console.

There is a useful difference between killing and quarantining an agent. Killing it can interrupt a production workflow, and that interruption has its own cost. Quarantine freezes the agent’s authority while people investigate and decide what happens next. In plain English: stop the agent from acting, then sort out the cause without assuming that stopping inference alone closed every door.

No single setting should be treated as a universal off switch. Your response depends on the agent’s architecture and the paths it can use. Map those paths before an incident, then decide whether the response should stop inference, block the action surface, or do both.

Conceptual diagram showing inference and action paths isolated separately or together

Managed identity is the quieter half

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

Give each agent a distinct, managed identity. Scope that identity to the operations the agent needs. Do not let an agent inherit a person’s broad permissions just because the workflow runs on that person’s behalf. Connection-level access is not operation-level authorization: permission to connect to a system does not mean permission to perform every action in it.

For example, an accounts-payable agent might need to read invoice details and prepare a payment for review. It does not automatically need permission to change vendor bank details and release payment. Those are separate operations with different consequences. Give them separate scopes and require the right approval where policy calls for it.

Microsoft’s guidance for managing agent identities describes ways to view and disable agent identities, manage permissions, and monitor sign-in and audit logs. The implementation details differ across platforms, but the principle is consistent: identify the agent, give it a narrow scope, and make its access revocable.

When identity is right, quarantine is a targeted incident control. When identity is broad or shared, quarantine becomes the brake you reach for after the agent has already been able to do too much.

Before the next inquiry, prepare the answers

An incident review does not end with “we stopped it.” A reviewer will want to know what happened before the stop, who or what had authority, and what evidence supports your account.

Ask yourself:

  • 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, through which response path?
  • What did it do before the control took effect?
  • Can we replay the action and authorization record?

Australia’s proposed reforms make this more than an engineering exercise. The government says it will legislate mandatory safety and security protections and a new regulator. The rapid review also recommends considering a “failure to prevent” offence. Those are proposals, not enacted laws. But they make the direction clear: the ability to explain an incident and show reasonable controls is becoming a governance concern, not just an operations concern. (Australian government announcement; The Guardian’s report.)

The EU AI Act also includes incident-reporting obligations for providers of general-purpose AI models with systemic risk. The Commission’s recent requests for more information about AI incidents are a further signal of scrutiny, not a formal finding of wrongdoing. The exact scope of enforcement around pre-market testing remains contested. See Article 55 of the AI Act and the reporting on the Commission’s requests.

A stop without an evidence trail leaves the same questions unanswered. LangGuard Arbiter evaluates actions before tool calls execute, while the audit ledger records policy decisions and responses. The control needs to stop the action and preserve the context needed to explain the stop.

Illustration connecting least-privilege identity, narrow operations, quarantine, and recorded evidence

A practical checklist for your deployment

Before your next agent goes into production, or before the next review, make sure you can:

  1. Inventory every agent and action surface. Record its 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 or incident workflow invokes it through the SIEM, ticketing system, MCP, API, or GUI.
  4. Choose the response in advance. Set criteria for stopping inference, disabling the action surface, or stopping both.
  5. Record the event. Capture which agent was stopped, who or what initiated the control, when it happened, what actions preceded it, and what policy applied.

The kill switch conversation is worth having. Just do not let it end at the button.

Know what the agent can reach. Give it only the authority it needs. Make sure you can stop that authority and show what happened before and after the stop.

Govern the action surface. Then prove what you stopped.

Sources