You can't pull the plug after it kills people. For example see [1], where the DOJ mistargeted a school with AI.

There are scenarios where kill -9 isn't going to happen in time. What if the team that is harming people with AI is different from the one that is monitoring the harm? What if no one is monitoring? What if the user is intentionally malicious?

And wiping out humanity doesn't necessarily mean shooting people either. Every trader involved in the '08 financial crisis was locally acting in their own interests. Those could have easily been AIs optimizing trading strategies too.

[1] https://futurism.com/artificial-intelligence/us-military-pen...