What you're describing might be intended. Your argument hinges on the assumption that oversight was supposed to stop this, but it could be they just have a higher level of acceptable collateral damage.

Either way, it's the responsibility of humans to design risk controls. When the AI makes a mistake it should be auditable how it came to that decision. The humans need proper training and regular exercises and ultimately to be held accountable.

I believe this achievable for those that want to achieve it.