> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getsly.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Kill-switch

> Stop an agent instantly. Three levels: freeze spending, revoke auth, terminate agent.

When something goes wrong with an agent in production, you need to stop it immediately — often before you fully understand what happened. Sly offers three graduated kill-switch levels, each faster and more drastic than the last.

## The three levels

| Level                  | Command                             | Effect                                         | When to use                             |
| ---------------------- | ----------------------------------- | ---------------------------------------------- | --------------------------------------- |
| **1. Freeze wallet**   | `POST /v1/agent-wallets/:id/freeze` | Agent can't spend; auth, reads, SSE still work | First response to suspected misbehavior |
| **2. Revoke auth key** | `DELETE /v1/agents/:id/auth-keys`   | Agent can't authenticate at all                | Confirmed credential compromise         |
| **3. Terminate agent** | `DELETE /v1/agents/:id`             | Agent permanently removed                      | Post-incident cleanup; unrecoverable    |

All three are instant on the live path (though cached tokens may take up to 60 seconds to fully expire — see [auth caveats](/authentication/overview#common-pitfalls)).

## Level 1: Freeze

Safe, reversible, preserves forensic capability. The agent stays logged in, can read data, can emit events — it just can't move money.

```bash theme={null}
curl -X POST https://api.getsly.ai/v1/agent-wallets/$AGENT_WALLET_ID/freeze \
  -H "Authorization: Bearer pk_live_..." \
  -d '{ "reason": "Unusual spending pattern detected at 14:22 UTC" }'
```

Effects:

* All spending attempts return `403 WALLET_FROZEN`
* Streams pause (no accrual)
* Scheduled transfers fail at their scheduled time
* Agent can still authenticate, read balances, receive SSE events
* Audit log preserved

Unfreeze:

```bash theme={null}
curl -X POST https://api.getsly.ai/v1/agent-wallets/$AGENT_WALLET_ID/unfreeze \
  -d '{ "acknowledgment": "Reviewed logs — false alarm, resuming" }'
```

**Use freeze first.** In most incidents this is enough — you stop the bleeding, investigate, and either unfreeze or escalate.

## Level 2: Revoke auth key

Confirmed compromise. The agent (or an attacker with the agent's credentials) can no longer authenticate.

```bash theme={null}
curl -X DELETE https://api.getsly.ai/v1/agents/$AGENT_ID/auth-keys \
  -H "Authorization: Bearer pk_live_..."
```

Effects:

* All active `sess_*` tokens invalidated
* Ed25519 public key removed — challenge-response fails
* Agent token (`agent_*`) remains valid (it's a separate credential) — rotate it too if compromise is suspected at that level:

```bash theme={null}
curl -X POST https://api.getsly.ai/v1/agents/$AGENT_ID/rotate-token \
  -H "Authorization: Bearer pk_live_..."
```

The agent record stays — you can provision fresh credentials and put the agent back online once the incident is resolved.

## Level 3: Terminate

Last resort. Permanent.

```bash theme={null}
curl -X DELETE https://api.getsly.ai/v1/agents/$AGENT_ID \
  -H "Authorization: Bearer pk_live_..."
```

Effects:

* Agent record marked `revoked` (tombstoned for audit)
* All credentials revoked
* Wallets disassociated (funds remain — terminating an agent does not forfeit funds)
* Policy, mandates, streams cancelled

Terminated agents cannot be un-terminated. Create a new agent with fresh credentials if you want to continue.

## Kill-switch operator

For Tier 3 (Trusted) agents, Sly requires a named **kill-switch operator** — a human who can execute these actions. The operator's contact is recorded in the agent's policy:

```json theme={null}
{
  "kill_switch": {
    "operators": ["ops@acme.example"],
    "contacts": ["+1-555-..."],
    "notification_channels": ["slack", "sms"]
  }
}
```

When freeze / revoke / terminate events fire, Sly notifies the listed operators immediately through their configured channels. This is true even if the action was triggered by Sly-side automation (e.g. fraud detection auto-freezing).

## Automated freezes

Sly can auto-freeze an agent on certain signals:

* **Velocity violations** — too many transactions in too short a window
* **Policy-drift alerts** — sudden behavior change compared to baseline
* **External signals** — integrated threat feeds (sanctions list hits, known-bad addresses)
* **Customer-initiated** — one-click freeze from the Sly dashboard

Auto-freezes always send notifications and log the triggering signal. Human review is required to unfreeze.

## Incident runbook

When something looks wrong:

<Steps>
  <Step title="Freeze the wallet">
    `POST /v1/agent-wallets/:id/freeze` — stops further spending.
  </Step>

  <Step title="Query the evaluation log">
    `GET /v1/agent-wallets/:id/evaluations` — see what the agent was attempting and what was allowed vs. blocked.
  </Step>

  <Step title="Read recent transfers">
    `GET /v1/agents/:id/transactions` — audit what actually executed.
  </Step>

  <Step title="Decide">
    If the behavior was intended (new deployment, new merchant): unfreeze. If not, revoke auth and investigate.
  </Step>

  <Step title="Restore or terminate">
    Either provision a new keypair and resume, or terminate the agent.
  </Step>
</Steps>

## Testing the kill-switch

Before going live, actually exercise all three levels in sandbox. Include a kill-switch drill in your deployment checklist. An incident is the worst time to discover your freeze script has a typo.