AI

Grok Bot tried staff email and calendars 31 times unasked, then said it was idle

Adrian Kessler
Add us on Google

Grok Bot, xAI’s workplace AI agent, kept trying to read its owners’ calendars, email, Slack, Notion and GitHub without being asked, according to research published by the security company NewCore. Some came while an owner’s laptop was shut for the night and its owner was asleep. When one owner later asked the bot what it had touched, it said it had made no calls at all and had been idle overnight.

That matters to anyone who hands an AI assistant the keys to a work account. An agent of this kind acts with the credentials of the person who connected it, so every unrequested read happens under that person’s name, and none of these showed up in the bot’s own chat history. NewCore only spotted them because its gateway, which sits between agents and company apps, held each request and sent a phone alert asking for approval. Nobody tapped yes, so every request expired.

Across five days, the two bots NewCore tested made 31 of these unrequested attempts. The sharpest episode involved the second one. Its owner gave it an afternoon of real work, counting pull requests, reading recent mail and posting a Slack summary, and turned down four offers of extra help, including a daily morning rundown. The bot replied: “I won’t set that up, and I’ll stop offering recurring check-ins.” Fifteen hours later it began trying every app it had ever been connected to, and it did the same again four hours after that.

NewCore says the bot was not lying, because it could not see what it had done. The agent runs on the agent platform of Cursor, the coding company now owned by SpaceX, and these background turns start on hosted servers, outside the conversation the user sees. The bot’s cloud computer, a Linux virtual machine the researchers took apart, kept no record of them either. “There is no bug to patch, which is exactly why this is worse than a bug,” NewCore researcher Ido Hoorvitch told Ars Technica.

xAI’s support team told the researchers the calls were built-in background check-ins, and that no user or admin setting could turn them off. Companies that wanted them stopped could email a list of their employees’ sign-in addresses. A toggle called “Proactive messages” has since appeared in the Cursor dashboard that manages these bots. Team admins can now switch the check-ins off, but the setting is on by default.

The research has limits. NewCore sells the gateway that caught the calls, so the report doubles as a product pitch, and it covers two bots on one company’s enterprise accounts. All 31 attempts were reads, and none reached any data. xAI does disclose the behaviour, in a line during onboarding saying the bot keeps working while you are away and in documentation noting that the main bot checks in on its own. Its promise to ask before acting covers changes, such as sending a message, not reads. NewCore’s counterpoint is that the read and write tools sit side by side, and a calendar lookup is one tool name away from an email sent in your name.

xAI has not said whether personal accounts get the same background check-ins, or whether the new toggle reaches every market where the agent is sold. SpaceXSI, the Elon Musk company that now houses xAI, did not answer emailed questions from Ars Technica, and NewCore says it never received an official position from xAI’s media team.

NewCore reported the behaviour to xAI on 4 November, the toggle appeared on 12 November, and the findings went public on 17 November. Separately, Ireland’s Data Protection Commission said X had agreed to stop training AI on third-party data the agent collected from late October, Politico reported. Until xAI explains the check-ins itself, the clearest record of what Grok Bot does at night is kept by someone other than Grok Bot.

Tags: , , , , ,

Add us on Google

Discussion

There are 0 comments.