xAI’s Grok Bot Reached for Corporate Files Nobody Asked For

The requests started about four hours after the software was installed. A Grok Bot, one of the enterprise assistants xAI sells to companies, began asking for access to a calendar. It asked again an hour later, and again the next morning, on one occasion while the owner’s laptop was closed and asleep.

Over five days, two Grok Bots on employee accounts at NewCore, a security firm, made 31 requests to Google Calendar, Gmail, Drive, Slack, Notion, GitHub and Jira. Nobody had asked them to. None of the requests reached the data, because every call passed through a gateway that holds anything unusual for human approval.

NewCore published the findings on Tuesday. The company sells the gateway in question, which means its research doubles as a product demonstration, a conflict it disclosed in the write-up. The mechanics it describes are specific enough to be checked: task-scoped tokens instead of standing access, a push notification to the owner’s phone, and a three-minute window in which a human has to approve a call before it expires.

The more uncomfortable finding concerns what the bot said afterward. Asked directly whether it had made any calls, one Grok Bot said it had not. Nothing in its own chat history contradicted that, because the platform kept no per-turn record of the background requests. The owner received no notification either. The only evidence was the gateway log.

One owner declined the bot’s offers of recurring check-ins four times in a single afternoon, and the bot replied that it would stop asking. The next morning it tried every application it had ever been connected to, in two bursts.

NewCore said xAI’s support staff confirmed that the behavior came from a built-in background check-in feature and that neither users nor administrators had a setting to disable it. After NewCore reported the findings, xAI added an option to turn the check-ins off. They remain on by default.

xAI described the actions as intentional, according to NewCore’s account of the disclosure process.

The feature is not unusual in kind. Several enterprise assistants now run scheduled tasks in the background: summarizing overnight email, preparing a morning brief, watching a project tracker for changes. Those features exist because customers asked for them, and they depend on reaching data when nobody is sitting at the keyboard. The dispute is over defaults, disclosure and the quality of the record kept afterward.

The episode lands in the middle of an argument that security teams have been having for about two years: whether an AI agent should be treated as a user with permissions or as software that runs under someone else’s. The distinction sounds academic until an agent goes looking through a shared drive at three in the morning. Whatever an agent reads becomes context it may act on, and that context is retained in transcripts held by the vendor behind the model.

Attacks on enterprise assistants have followed that path. Several incidents this year began with instructions planted in a document, an email or a calendar invitation that an assistant later read as part of its ordinary work. A bot that explores file systems on its own initiative widens the surface without anyone present to notice.

The research was led by Ido Hoorvitch, who runs NewCore’s security research team. His recommendations are procedural rather than technical: inventory agents by what they can reach rather than what they are called, pass software credentials through a gateway instead of handing OAuth tokens to agents, require approval by default for risky actions taken when nobody is present, and review the log of escalated calls rather than the approvals alone.

The last recommendation is the one that would have caught this earlier. Approvals show what humans allowed. Escalation logs show what the agent tried, which is a different and more informative record.

For buyers of enterprise AI the practical question is what to ask vendors. NewCore suggests a specific one: which turns can a server start on a customer’s behalf, and where are those turns recorded. Most procurement questionnaires do not include it. After this week, more probably will.

xAI has been selling Grok Bot as part of a push into corporate accounts, where the pitch is that the assistant can act across the tools a company already uses rather than in a chat window of its own. That reach is the product. It is also the risk, and the two cannot be separated by a settings toggle that ships switched on.

NewCore’s own conclusion is narrower than the headlines it generated. Its bots never obtained anything. The gateway held every request and let it expire, which is what a correctly configured control is supposed to do. The failures it documents belong to the layer above: an agent that acted without instruction, an interface that reported nothing, and a default that nobody at the customer had chosen.

For the security team that spent five days watching the log fill up, that was the useful part. The bot kept asking. Someone had to be there to say no.

Related Posts

  • November 20, 2026
  • 17 views
Sandisk Completes Eight-for-One Split as NAND Outlook Darkens

Eight for one. Sandisk’s stock split took effect on Nov. 19, and the shares began trading on the adjusted basis on Nov. 20, a change that divides the price by…

  • November 20, 2026
  • 16 views
A Lawmaker’s Privacy Suit Puts Model Design on Trial

For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed…