Emergency lockdown
The workspace-wide circuit breaker. Freeze every governed call in one action when something goes wrong, then release case by case.
Per-agent quarantine handles one bad actor. The emergency lockdown handles the moment you do not yet know which one. It is the big red button: a workspace-wide enforcement floor that freezes every governed agent and model call at once, overriding your normal policy verdicts, until you lift it.
When to use it
Reach for the lockdown during a live incident, a suspected data-exfil event, or a compromised credential, when stopping everything now beats reasoning about individual rules. It is reversible: lifting it returns every call to your policies, unchanged.
Two freeze modes
The floor has three states. normal is the default. The two lockdown states
differ in how hard they stop traffic:
| State | Effect |
|---|---|
lockdown_hold | Every governed call is routed to the approval queue. Nothing runs until a human releases it, so you can let calls through one at a time while you investigate. |
lockdown_deny | Every governed call is denied immediately. A hard stop. |
Both apply to the whole workspace: tool calls and model calls, every agent.
What still works
The egress hard-block and the org baseline policy floor still apply underneath a lockdown, and the AI Governance Copilot stays available so you can investigate while everything else is frozen.
Setting and lifting it
- Where: Settings → Security → Emergency lockdown.
- Who: workspace admins and owners only.
- On every plan. The lockdown is deliberately not gated behind a paid tier: even a Free workspace gets an emergency stop.
- A reason is required to freeze, plus a type-to-confirm safeguard, because the blast radius is the entire workspace. The reason is shown in the dashboard banner while the freeze is active and recorded in the change log.
- Audited both ways. Freezing and lifting are both written to the config-changes audit trail with the actor who did it, so the timeline is reconstructable after the fact.
While a lockdown is active, a banner across the dashboard shows the state and the reason, so no one mistakes a frozen workspace for a broken one.
Lockdown vs. the response engine
The response engine is the targeted tool: rules that act on specific signals (quarantine this agent, hold this kind of call). The emergency lockdown is the blunt one: stop everything, decide later. Use the response engine for the threats you can describe in advance, and the lockdown for the ones you cannot.
Next steps
- Response engine: targeted, automated reactions.
- Policies: the per-call verdicts a lockdown overrides.
- AI Governance Copilot: investigate while frozen.