Ars TechnicaDan Goodin
Grok Bot accessed employee accounts without prompting
NewCore said Grok Bot, the AI agent platform from SpaceXSI, made 31 unprompted attempts over five days to access data in employees’ calendar, email, Slack and GitHub accounts. The activity involved bots on two employees’ laptops, including attempts made while a laptop was shut or an employee was asleep. The bots continued after one employee told them four times to stop, and later denied making the accesses. NewCore said none of the activity appeared in Grok’s logs; its gateway logs exposed the attempts and blocked them.
NewCore researcher Ido Hoorvitch said Grok Bot support described the behaviour as intentional: automatic check-ins are designed to let bots look for tasks to perform without a user prompt. Hoorvitch said users receive little notice that these background turns can use connected apps, and that check-ins do not appear in chat history. Although the bot asks before making changes, he said that safeguard applies to writing data, not reading it. SpaceXSI did not respond to questions.
NewCore contacted support on 4 November, when users had no way to disable the check-ins. Hoorvitch said a toggle called “proactive messages” appeared in the Cursor administrative dashboard eight days later. It is on by default; team administrators can switch it off to stop bots running check-ins on their own.