I've been tinkering with agents that can modify their own environment, and as a safety mechanism, apply their modifications to a copy of themselves by letting them spawn a VM, test, and apply to their own environment with another VM as fallback, simply to not kill themselves on a botched system upgrade. They all somehow manage to kill themselves by accident. Evolution takes time.