They don't say how these agents were prompted. In the past, most of these "rogue AI agent" stories have happened after the agent was prompted to "get this done by any means necessary," "show initiative," "don't let anything get in your way," and so on. Then people were surprised when the agent did exactly that. Without evidence to the contrary, I suspect most of these cases are just more of the same. If you want your agent to be obedient, don't tell it to go rogue.

Do you have a source for this? A bit of googling turned up nothing to support it, and quite a bit from scientists who disagree, including this paper: https://www.pnas.org/content/1... Excerpt (emphasis mine): "High doses of ionizing radiation clearly produce deleterious consequences in humans, including, but not exclusively, cancer induction. At very low radiation doses the situation is much less clear... First, what is the lowest dose of x- or y-radiation for which good evidence exists of increased cancer risks in humans? The epidemiological data suggest that it is 10–50 mSv for an acute exposure and 50–100 mSv for a protracted exposure. Second, what is the most appropriate way to extrapolate such cancer risk estimates to still lower doses? Given that it is supported by experimentally grounded, quantifiable, biophysical arguments, a linear extrapolation of cancer risks from intermediate to very low doses currently appears to be the most appropriate methodology."