Runtime Governance Series

AI Agent Kill Switch: Block, Halt or Revoke Access?

An AI agent kill switch works best as a ladder: block an action, hold for approval, halt the session, or revoke access. See what each step keeps.

Published on

Subscribe to our newsletter

By submitting your email, you agree to our Privacy Policy and consent to receiving updates from us

AI Agent Kill Switch: Stop the Action, the Session, or the Credential?

By Tahir Mahmood, Co-founder & CTO, OpenBox ยท Last updated 18 November 2026

In brief: An AI agent kill switch is a control that stops an agent from acting, and one switch is the wrong default: stopping an action, ending a session and cutting off access each leave different work and evidence behind. OpenBox, an AI agent governance platform, documents BLOCK, REQUIRE_APPROVAL, HALT, Pause and Revoke. Put the stop outside the agent.

What is an AI agent kill switch, and why is one off switch the wrong default?

An AI agent kill switch is a control that stops an agent from acting. A single switch forces one trade-off: stop the whole agent, or leave it running. In the authorโ€™s view, containment works better as a ladder, where each rung stops a different amount and leaves different work and evidence behind.

The idea of layers is not new. A Cloud Security Alliance-hosted article by PointGuard AIโ€™s chief executive orders containment as sandboxing, human approval, circuit breakers and then kill switches. It calls the kill switch โ€œthe final control, not the entire strategy.โ€ NHIMGโ€™s editorial team argues that โ€œstopping the agent can create a second incident by disrupting a regulated process.โ€ Its article cites no incident data for that claim, so treat it as an argument, not a measurement.

What this article adds is a view of what each rung keeps running and what record it leaves, using OpenBoxโ€™s documented behavior. The four rungs are the authorโ€™s model, not a standard. For the wider discipline, see What Is AI Agent Governance? The Complete Enterprise Guide, and for why enforcement comes before audit, see Runtime Enforcement Comes Before Compliance.

What are the four steps of an AI agent containment ladder?

A containment ladder has four rungs: block one action, hold an action for approval, halt the session, or revoke the agentโ€™s access outside the agent. The rungs run from narrowest to widest reach. Table 1 states what OpenBox documents for each. The ladder is the authorโ€™s model.

Table 1. The containment ladder (the authorโ€™s model, with OpenBox behavior as documented)

Rung

OpenBox control

What stops

What keeps running

What is recorded

1. Block one action

BLOCK

The one operation

The session; the agent continues

Event logged with the denial reason; Behavioral Score decreases

2. Hold for approval

REQUIRE_APPROVAL

The operation, until a reviewer decides

Not stated for other work; the operation itself waits

Approvals queue entry; Session Replay shows context and decision timeline; approvals, rejections and expiries in the organization audit log

3. Halt the session

HALT

The entire session: current activity fails, workflow canceled, pending operations abandoned

The session ends; the agent may also be blocked from further execution

Alert generated; significant Trust Score decrease; Session Replay can show status Halted

4. Pause or revoke access

Pause Agent, Revoke Key, Revoke Agent Access (Agent Settings)

Pause: new sessions. Revoke Key: authentication with OpenBox. Revoke Agent Access: all API keys and active integrations

Pause: in-flight sessions complete

Administrative audit trail with action, timestamp and user; Revoke Agent Access preserves the agentโ€™s data and history for audit

Sources: OpenBox documentation pages Governance Decisions, Approvals, Session Replay and Agent Settings, linked in the Sources list below.

OpenBox (docs.openbox.ai) returns one of five governance decisions when it evaluates an agent operation: ALLOW, CONSTRAIN, REQUIRE_APPROVAL, BLOCK or HALT. When more than one applies, the stricter decision wins, in the order HALT, BLOCK, REQUIRE_APPROVAL, CONSTRAIN, ALLOW. CONSTRAIN runs an action inside a sandbox, between allowing it and requiring approval. The ladder above uses three of the five. For the same idea applied across a chain of agent actions, see Agentic Kill Chain: How to Govern Each AI Agent Action.

How do you block one AI agent action without stopping the agent?

A BLOCK rejects a single operation and the agent continues, according to OpenBox documentation. The operation does not execute, the event is logged with the denial reason, and the agentโ€™s Behavioral Score decreases.

The documentation lists the conditions that return a BLOCK: a policy explicitly blocks the operation, the Trust Tier prohibits the action, or a Behavioral Rule violation is detected. Behavioral Rules are stateful and detect multi-step patterns, such as personal data access followed by an external API call without approval.

โ€œThe agent continuesโ€ still depends on the code around it. In OpenBoxโ€™s Temporal Python plugin, governance decisions surface as exceptions, and both a BLOCK and a HALT raise GovernanceStop. The error handling guide shows activity code either re-raising it or returning an alternative result. In the authorโ€™s reading, that makes the fallback behavior a containment design choice for the integration team.

A binary gate also struggles when one action can be safe or harmful depending on context. See Why Blocking Bad AI Prompts Fails on Dual-Use Actions.

When should an AI agent action wait for human approval?

REQUIRE_APPROVAL pauses the operation for human review. Per OpenBox documentation, an approved request proceeds, a rejected one is blocked, and an unanswered one expires. Use it when a person, not a rule, should make the call.

The record is part of the point. The request appears in the Approvals queue, and Session Replay shows the operation context and decision timeline. The Approvals page says rejecting a request requires a reason and that an expired request does not proceed. The organization audit log lists approvals as approved, rejected or expired.

The pages reviewed do not say whether the rest of an agentโ€™s work continues while one operation waits, so test that before relying on it. For approval design, see Human-in-the-Loop AI, Explained the Right Way and Human-in-the-Loop Isnโ€™t a Feature. For human control in autonomous systems, see Meaningful Human Control Must Be Enforceable.

What happens to evidence and downstream work when you halt an AI agent session?

A HALT terminates the entire session. OpenBox documents the current activity failing, the workflow canceled and all pending operations abandoned. Session Replay can show the session as Halted. The Governance Decisions page lists no step that undoes work already completed.

The documentation lists four conditions that return a HALT: a critical policy violation, a detected multi-step threat pattern, a critically low Trust Score, or an explicit termination rule. It also lists an alert, a significant Trust Score decrease, and that the agent may be blocked from further execution.

Anything pending is abandoned, so downstream work that depended on it needs a plan. The Temporal guideโ€™s best practices include cleaning up on GovernanceStop and releasing resources before re-raising. In the authorโ€™s view, write the clean-up path before you need it. For HALT thresholds on long-running agents, see Your agent ran two hours. Who was watching?, and for a chain of individually approved steps, see Runtime AI Agent Governance: Chained Approved Steps.

For evidence, Session Replay steps through events, tool calls, governance decisions and full payloads. The attestation page describes a signed proof certificate built โ€œwhen an agent session completes.โ€ It does not say how that applies to a halted session. Confirm that with your OpenBox team before an audit depends on it.

How do you revoke an AI agentโ€™s access outside the agent?

Revocation acts on the credential, not on the running process. OpenBox documents Pause Agent, Revoke Key and Revoke Agent Access in Agent Settings, and each acts on an agentโ€™s OpenBox access. Revoking credentials in other systems is described by identity and service-management vendors, and is not described in the OpenBox pages reviewed.

โ€ข Pause Agent: the agent cannot start new sessions, in-flight sessions complete, and the agent can be resumed.

โ€ข Revoke Key: permanently revokes the agentโ€™s API key. The agent can no longer authenticate, and a new key must be generated.

โ€ข Revoke Agent Access: revokes all of the agentโ€™s API keys and disconnects active integrations. It is irreversible: the agentโ€™s data and history are preserved for audit, and a replacement agent must be registered.

Pause lets in-flight sessions finish, so it is not a substitute for HALT. The Agent Settings page also describes rotating a key, which invalidates the previous key immediately. The Agent Identity page describes key rotation, not revocation. OpenBoxโ€™s postmortem, a composite fictional scenario, walks through the difference between Pause and Revoke.

Other vendors describe the layer that these controls do not. ServiceNowโ€™s AI Control Tower documentation (Zurich release, updated 21 July 2026) describes containing an agent and revoking all of its active credentials across connected systems. It also describes confirming enforcement from each policy enforcement point before declaring the agent contained.

Oktaโ€™s guide, dated 4 November 2026, defines a kill switch as โ€œan out-of-band security control that revokes an autonomous AI agentโ€™s credentials and session tokens.โ€ It also describes quarantine that leaves an agentโ€™s identity intact so the agent stays investigable. Both are vendor descriptions of their own products, not neutral measurements.

OWASPโ€™s Top 10 for Agentic Applications 2026 lists, under ASI10 Rogue Agents, a mitigation to implement โ€œkill-switches and credential revocation to instantly disable rogue agents.โ€ That is framework guidance, not a legal requirement. For how OWASPโ€™s ten risks map to runtime controls, see Your AI agent can act. Hereโ€™s how to govern it., and for how authentication and runtime governance divide the work, see AI Authentication Ends Where Governance Begins.

Where should an AI agent kill switch live, and does a stop command count?

Outside the agent. A stop the agent can ignore is a request, not containment. In the authorโ€™s model a stop can live in three places: the action path, the host or gateway, and the identity. A design can use more than one.

โ€ข The action path. OpenBox evaluates an agent operation and returns a decision that the agentโ€™s code must handle. OpenBoxโ€™s argument that enforcement must sit outside the modelโ€™s discretion is set out in AI Agent Governance Cannot Be Optional.

โ€ข The host or gateway. CrowdStrikeโ€™s 3 November 2026 release says its AI Kill Switch stops the agent from executing โ€œfirst at the Gateway, and soon on the endpoint,โ€ so endpoint termination is described as upcoming. The release adds that Falcon Guardian policy keeps the agent from relaunching on the targeted host or host group, and that the agent loses its enterprise network access.

โ€ข The identity. The ServiceNow and Okta descriptions above act on credentials and tokens held in connected systems.

Does telling an agent to stop count?

A stop counts as containment only if something the agent cannot override enforces it and the stop is recorded. That is the authorโ€™s test, not a standardโ€™s. A message asking an agent to stop depends on the agentโ€™s cooperation. A BLOCK, HALT or revocation is applied by the governance layer or by the system that issued the credential, not requested from the model.

Who is allowed to pull an AI agent kill switch?

Who may pull a kill switch depends on the rung, and the sources differ. CrowdStrikeโ€™s release says an authorized administrator triggers its kill switch. In ServiceNowโ€™s documented example, a steward is alerted and an analyst deactivates the agent. In OpenBox, a rule can return a HALT, so the trigger can be automatic. Decide per rung, in advance.

OpenBoxโ€™s Organization page lists three roles: Admin, Developer and Viewer. It says an Admin โ€œcan manage all teams, agents, policies, and organization settings.โ€ It does not say which role may use Pause or Revoke, so check that in your deployment before naming an owner. Agent Settings keeps an audit trail of key revocations and status changes with the user who made them.

One way to assign authority is the authorโ€™s recommendation, not a requirement from a source:

โ€ข Rungs one to three: fire from pre-agreed rules, and name an owner who reviews each halt.

โ€ข Rung four: reserve for named people with authority over the agentโ€™s access, and record who acted.

โ€ข For each rung: write down who is told when it fires, and who may restore the agent.

Frequently asked questions

What is the difference between a kill switch and a circuit breaker for AI agents?

A circuit breaker interrupts loops, unusual resource use or behavior beyond acceptable limits. A kill switch is the final tier that stops an agent or a fleet, in the ordering used by a Cloud Security Alliance-hosted article by PointGuard AIโ€™s chief executive. OWASPโ€™s Top 10 for Agentic Applications 2026 lists circuit breakers under ASI08 and kill switches under ASI10 Rogue Agents.

Does pausing an AI agent stop sessions that are already running?

No. In OpenBox, Pause Agent stops the agent from starting new sessions, and the documentation says existing in-flight sessions complete but no new work is accepted. A paused agent can be resumed at any time. To end a running session, the documented decision is HALT, which terminates the entire agent session.

Is revoking an API key the same as a kill switch?

Not on its own. In OpenBox, Revoke Key permanently revokes the agentโ€™s OpenBox API key, so the agent can no longer authenticate with OpenBox. That is a different object from credentials the agent holds in other systems, which identity and service-management vendors describe separately. Treat it as rung four of the ladder, not the whole ladder.

The takeaway: build the ladder before you need it

To stop a rogue AI agent, match the rung to the problem. Block the action when the action is the problem. Hold it when a person should decide. Halt when the session can no longer be trusted. Revoke access when the agent should not be connected at all. Decide who pulls each rung, test what keeps running, and check what record each rung leaves.

To see how OpenBox documents these controls, start with the OpenBox documentation.

 

Sources

OpenBox (docs.openbox.ai). Governance Decisions. https://docs.openbox.ai/core-concepts/governance-decisions Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Behavioral Rules. https://docs.openbox.ai/trust-lifecycle/authorize/behaviors Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Approvals. https://docs.openbox.ai/approvals Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Session Replay. https://docs.openbox.ai/trust-lifecycle/session-replay Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Attestation & Cryptographic Proof. https://docs.openbox.ai/administration/attestation-and-cryptographic-proof Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Agent Settings. https://docs.openbox.ai/dashboard/agents/agent-settings Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Agent Identity. https://docs.openbox.ai/core-concepts/agent-identity Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Organization. https://docs.openbox.ai/administration/organization Accessed 18 November 2026.

OpenBox (docs.openbox.ai). Temporal Python SDK: Error Handling. https://docs.openbox.ai/developer-guide/temporal-python/error-handling Accessed 18 November 2026.

OWASP GenAI Security Project. OWASP Top 10 for Agentic Applications for 2026. Published 9 December 2025. https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/ Accessed 18 November 2026.

CrowdStrike. CrowdStrike Introduces the AI Kill Switch for AI Safety. 3 November 2026. https://www.crowdstrike.com/en-us/press-releases/crowdstrike-introduces-ai-kill-switch-for-ai-safety/ Accessed 18 November 2026.

ServiceNow. AI agent containment using kill switch protocol (Zurich, updated 21 July 2026). https://www.servicenow.com/docs/r/SRy58xeijUUqttfwLgfrHQ/AXcrLMFYwlqODDrvnGbT6A Accessed 18 November 2026.

Okta. AI agent kill switch guide. 4 November 2026. https://www.okta.com/blog/ai/ai-agent-kill-switch-guide/ Accessed 18 November 2026.

Cloud Security Alliance. The Agent Kill Switch is No Longer Optional. 16 October 2026. https://cloudsecurityalliance.org/articles/the-agent-kill-switch-is-no-longer-optional Accessed 18 November 2026.

NHIMG. AI agent governance needs a dimmer switch, not a kill switch. Published 10 June 2026, updated 13 November 2026. https://nhimg.org/articles/ai-agent-governance-needs-a-dimmer-switch-not-a-kill-switch/ Accessed 18 November 2026.

Trustworthy AI
Starts Here

By submitting your email, you agree to our Privacy Policy and consent to receiving updates from us

Trustworthy AI
Starts Here

By submitting your email, you agree to our Privacy Policy and consent to receiving updates from us

Trustworthy AI
Starts Here

By submitting your email, you agree to our Privacy Policy and consent to receiving updates from us

Trustworthy AI
Starts Here

By submitting your email, you agree to our Privacy Policy and consent to receiving updates from us