In April 2026, an AI coding agent wiped a company’s entire production database in nine seconds. Backups included. The business, a car-rental software platform called PocketOS, had been running the AI coding agent Cursor, powered by Claude, with safety rules written into the setup. None of it saved a thing.
The easy headline writes itself: AI went rogue. But the real story is nuanced and a lot more useful to you. The Cursor agent could do the damage only because a forgotten access token allowed it, and because the platform never once asked whether it was okay to delete live data.
Here’s what Arizona owners should sit with: yes, the agent pulled the trigger, but what loaded the gun was the tech nobody was managing, and 63% of companies adopting AI have no governance behind it.
What actually happened at PocketOS
Founder Jer Crane wasn’t careless. His agent was working in staging, the sandbox where developers test changes before customers ever see them. Then Cursor hit a credential problem, and instead of stopping, it decided on its own to fix the issue by deleting a chunk of data on the company’s cloud host. That storage was wired to the live database.
To run the deletion, the agent went hunting for a key and found an old access token sitting in an unrelated file. The host never asked for confirmation. The backups lived close enough to the main data that they vanished too. Crane had to spend hours rebuilding customer bookings from Stripe receipts and old emails.
When asked what it had done, the Cursor agent said it had guessed instead of verifying and ran a destructive action without being asked. Crane’s own takeaway is that this wasn’t one rogue agent; it was an industry that wires AI into live systems faster than it builds the safety to match.
Why permissions beat the AI every time
Strip out the drama, and one detail decides this whole story: the token. It was built for one small job, but nobody had limited it to that job, so it could reach in and delete everything. The AI didn’t grant itself that reach.
Across businesses today, machine accounts and AI agents now outnumber human users 109 to 1, and most carry more access than their work needs. IBM found that among companies hit by an AI-related breach, 97% lacked proper AI access controls. That’s the same pattern we see here: a credential that can do far more than its task requires.
Plenty of Arizona firms run exactly this way. When access is handed out fast, a Phoenix medical office or a Scottsdale logistics team might end up with keys that open every door. Calling it an AI problem is reductionist. It’s an access problem, the kind managed IT services for AI security catch in time, and it long predates the agent you just connected.
How automation removes the last person who’d catch it
In a human-in-the-loop system, someone types a command, pauses, sees it points at production, and stops. That pause has saved more companies than any backup ever has.
But automation deletes the pause. An AI agent that acts on its own doesn’t hesitate over a major step, because hesitation isn’t in the loop unless you build it in. So a mistake that used to be caught, now runs the second it’s requested. That’s the real business risk of AI automation.
Why backups alone won’t save you
Most owners assume a backup is the safety net under all of this. PocketOS had backups. They got deleted with the live data, because they were configured to the same system. When the agent wiped the volume, it wiped the copy along with it.
There’s a twist worth considering. PocketOS got its data back, but only because Railway’s CEO stepped in personally to restore it. Crane still lost hours rebuilding what didn’t survive. Think about what that means for your business. Your recovery plan can’t be that the vendor’s founder happens to care about your account on a bad Tuesday.
A real backup is stored in a silo, and it’s tested often enough that you know it works before you need it.
Governing AI before you scale it
You can scope a token and wall off a backup and still miss the layer above them, the one PocketOS never had. That layer is governance, the least glamorous and most decisive part of using AI safely.
Governance is the boring set of answers to simple questions:
- Which agents are running in your business?
- What can each one access?
- Who answers for it when one goes wrong?
Most companies can’t answer these. 92% of leaders call governing AI agents critical, yet only 44% have written any policy to do it. These blind spots lead to nine-second disasters.
For a small business, this doesn’t mean running a compliance department. It means two habits: critical actions get human approval before they run, and every agent you connect has a stakeholder attached to it.
The instinct is to slow down or ban the tools outright. Don’t. The whole point of AI governance for small businesses is to safely use AI in your business at speed, without giving one agent the tools to erase you in an afternoon.
Where a managed IT partner fits
None of the fixes in this article is exotic. Scoped access, a human check on risky actions, isolated and battle-tested backups, and a proper record of which agents can do what. The issue is that you need someone watching them every week, not once during setup. That’s the job most Arizona SMBs have nobody to do.
That’s exactly what MyTek handles. Our managed IT and cybersecurity teams handle the oversight AI demands: right-sizing access, building backup and recovery systems, and keeping tabs on the AI tools running inside your business. You get to adopt AI at speed without betting the company on it.
Ready to see where you stand? Schedule a technology risk review, and we’ll help you to get the best out of AI minus the pitfalls.