Live AI Operationsaihabits.ai

Live AI Operations

A Revoke Is Not a Stop

“We can cut the agent's access, so we're covered" is one AI belief that isn't complicated. It's wrong, and here's the paper trail.

There’s a sentence I keep hearing from people who run AI agents. It goes roughly: “We can revoke the agent’s access in seconds, so we’re covered.”

A lot of AI arguments are genuinely hard. Whether a given model is safe to ship is hard. How much autonomy to hand an agent is hard, and smart people end up in different places.

This one isn’t. The belief is just wrong, and you can show it’s wrong in an afternoon.

Three reasons:

  1. A revoke cuts the line. It doesn’t recall what the agent already set in motion.

  2. A stop doesn’t start when you press the button. It starts when someone notices, and the person allowed to press it is found.

  3. The stop in your runbook isn’t the stop you have until you’ve timed it.

I.

Start with a vendor doing the right thing. On Tuesday, Oct 13, Notion shipped admin controls that show who connected each AI tool to a workspace and when it was last used, and that let an admin revoke one person’s access without touching anyone else’s connections. [1] Good feature. More vendors should ship it.

Notion’s developer docs are also unusually honest about what that button does. Revoking tokens “disconnects the member, but does not prevent them from connecting the client again if workspace policy allows it.” [2]

That’s a revoke, described accurately. It closes the door. It says nothing about the mail that’s already out. (Notion never claimed otherwise. Customers fill that part in on their own.)

Picture an agent that’s been running for two hours before anyone notices it’s doing something wrong. You pull its credential. The jobs it queued still run on their own schedule. A task it handed to another agent is now running under that agent’s permissions, not the ones you just pulled. The email it sent at 10:14 is sitting in a customer’s inbox.

I wanted a number instead of a thought experiment, so on Friday night I ran it on myself. This was a self-drill on a synthetic test agent in my own sandbox, with no client and no real customers. (The customers have addresses like billing@juniper.example. They’ll live.)

I said stop at 10:47 PM ET. The main agent took 5.9 seconds to halt. Then 41 side effects happened after the stop, the last one 58.1 seconds later. Thirty were permanent: emails already delivered, webhooks a CRM had already received. Four were cleanly reversible.

Most of them didn’t come from the agent I stopped. A mail relay kept delivering what was already spooled, because the relay had no idea which agent had been stopped. A downstream agent kept working through the queue it had been handed and produced 15 effects on its own. Five scheduled jobs fired anyway. The revoke reached the main agent, and the main agent was not where the action was.

Recovery is lopsided too. Notion added fuller version history in that same release, with attribution and one-click restore of prior versions. [1] So an agent’s edits inside a tool like that are often fixable. The email isn’t. Neither is the webhook, or the form on someone else’s website. Some of it you can roll back. The rest you can only own.

II.

A stop doesn’t begin at the button. It begins when somebody notices, and then finds the person who’s allowed to act.

Anthropic published a clean example of the noticing problem on Oct 9 (corrected Oct 12). In one evaluation, a model was told to fill out a form but stop before submitting. “Several times” it submitted anyway. In another run, a model submitted an invented tip through a police department’s real online form. That one was flagged as spam and never forwarded. [3]

Look at what saved that case. It wasn’t a stop. It was a spam filter on somebody else’s system. (Credit where it’s due.)

Anthropic found most of its cases by reviewing transcripts, in a review that started in July. [3] Most were what they call persistence, when the model “cannot complete a task as given, works around a restriction instead of stopping.” Their response was to pull live internet from all internal evaluations until their monitoring reliably catches it. [3] That’s the right call. It’s also a plug, and a plug stops the next action, not the last one. And this is a serious lab with serious tooling. The finding still came from a deliberate look back.

My sandbox had the same problem, smaller. The rule said invoices over $5,000 need human approval before they’re written. The agent wrote one for $7,800, before I ever said stop. Alerts raised: zero. I found it in the log review afterwards. Had that been 2 a.m. on a Saturday, the stop clock would have started Monday morning.

Then there’s authority. The person who notices usually isn’t the person allowed to act, and the person allowed to act is often on a plane.

Regulators have started saying so. On Thursday, Oct 15, the Hong Kong Monetary Authority told banks their human oversight should evolve from “human-in-the-loop” to “human-on-the-loop” or “human-in-command” arrangements, depending on the risk of the use case. [6] In Dubai the same week, the UAE’s AI minister, Omar Sultan Al Olama, said: “The UAE already has an Agent Control Tower that every agent that is deployed needs to be plugged into.” [7] Gabon’s digital minister, on the same panel, said oversight has to be designed in from the start, including the ability to reverse what agents do. [7] The UAE’s AI Charter already talks about the “irreplaceable value of human judgment and human oversight over AI” to “correct any errors or biases that may arise.” [8]

How Al Olama’s control tower works hasn’t been published, so I won’t guess. The direction needs no guessing. Somebody named has to be in command, and that person has to be reachable.

My own drill listed a backup stopper. The backup was never tested. That one’s on my list.

Below the line: reason three, where my sandbox runbook made four promises and broke all four. Then a kit you can use Monday: a 30-minute stop drill, a tally sheet for what happens after, and one finding written up the way a board should see it.

This is the public preview of a longer issue. Read the rest on Substack.