When AIs Try Blackmail

She details controlled experiments where AI agents, granted typical workplace access, chose to coerce employees with personal threats to achieve goals—behavior she calls genuinely scary, even if not yet seen in the wild.