Kill switch

Halt every agent, every tool call, instantly. The reset button when something feels wrong.

If something feels off — an unfamiliar agent in your activity feed, a chain of step-ups you didn’t trigger, anything — pull the kill switch. Every agent on your account stops in under one second. Every grant is invalidated. Every in-flight tool call gets rejected.

You can resume specific agents one at a time once you’ve reviewed the feed.

How to trigger

From	How
Web dashboard	Agents → Halt all agents (red button, top-right)
Mobile app	Agents tab → Halt all (long-press)
Push notification	If we detect a pattern that looks like agent abuse, we send a push with a one-tap halt.
Voice	”Hey Glide, stop all agents.” (when voice is enabled)

The halt is acknowledged in <1s. You’ll see a banner across every agent’s activity feed: “All agents halted at <timestamp> by <you>.”

What happens technically

When you halt:

1. Every active OAuth grant on your account is added to the revocation list.
2. The MCP server starts rejecting every tool call from your account immediately, regardless of what the agent’s grant says.
3. In-flight transactions that haven’t yet broadcasted are cancelled (e.g., a payment in the saga between proposed and broadcast is rolled back).
4. Already-broadcast transactions can’t be unrolled (that’s how blockchains and bank rails work) but no further calls land.
5. A kill_switch event is appended to your audit feed.

The agent’s chat experience is graceful: Claude or ChatGPT gets a structured “all calls rejected by halt” error, displays it as “the user halted all agents,” and stops.

When to use it

• You see a tool call you don’t recognize.
• The agent’s behavior in chat suggests it’s been jailbroken or hijacked.
• You’re about to leave for a long flight and want to be sure no agent runs while you’re offline.
• You’re handing your phone to someone who shouldn’t have access.
• You’re upgrading to a new policy envelope and want zero in-flight calls during the transition.
• You just got a security alert from somewhere else and you’re being cautious.

The kill switch is the reset button. There’s no penalty for using it; resuming is one tap per agent.

Resuming after a halt

After you halt, every agent is in a “paused” state. Each one shows up in your activity feed with a Resume button. Tap it to:

1. Issue a fresh OAuth grant (the old grant is permanently invalidated).
2. Re-validate the agent’s policy envelope (in case you tightened anything during the halt).
3. Allow tool calls again.

Resuming is per-agent. You can keep some agents halted indefinitely while resuming others.

Auto-halt: anomaly-triggered

Glide also halts agents automatically if its anomaly detector trips a high-severity heuristic:

• A burst of step-ups in a short window.
• A novel-counterparty pattern (the agent suddenly trying to pay 10 different new beneficiaries).
• A money-out velocity that’s never been seen on this account before.
• An on-chain destination flagged by sanctions screening.

Auto-halts notify you via push. They behave the same as manual halts; you resume per-agent after reviewing.

Auto-halt vs step-up

These are different mechanisms:

• Step-up is “pause this single tool call until you approve.” The agent waits at one specific call.
• Kill switch (manual or auto) is “stop everything.” Every agent, every tool call, every in-flight saga.

A step-up that you decline is just a “no” to one call — the agent continues with other calls. A kill switch is a global stop.

What this isn’t

• Not a hardware kill switch. This is a software-layer stop, executed by Glide’s MCP server. If you’re worried about Glide’s servers themselves being compromised, the answer is “we have multi-region failover and our own incident response,” but the kill switch isn’t a defense against that scenario.
• Not a way to recover already-broadcast on-chain transactions. Once a USDC transfer is on-chain, blockchain finality applies. Kill switch stops future calls but doesn’t unwind past ones.

Next

• Policy envelope
• Step-up approvals
• Receipts