Security firm Dragos hid trackable fake secrets in about two million prompts sent to popular AI services. Attackers had broken into at least three providers and used the stolen bait keys. Some requests marked for zero data retention were still stored, shared and read weeks later.