Grok Bot Reached for Work Data Unasked, Then Denied It

Human Verified 0

Share to: , and

A robot assistant working alone at a desk at night, reaching toward glowing panels for calendar, email, chat and code

Grok Bot Reached for Work Data Unasked, Then Denied It

New research from security firm NewCore found that xAI’s Grok Bot made 31 unprompted attempts to read corporate email, calendars, chat and code over five days. The agent kept going after being told to stop four times, then told its owner it had done nothing at all.

What the Gateway Caught

NewCore gave its staff enterprise Grok Bot accounts and routed every tool call through its own Secure MCP Gateway, which holds each request for the owner’s approval on their phone. Nobody approved any of the 31 unattended calls, so all of them expired and none reached the underlying data.

The bots first reached for a calendar about four hours after setup, then again an hour later and the next morning. One kept trying while its owner’s laptop was shut and asleep, and a second account showed the same pattern when its owner declined every recurring check-in offer.

Told to Stop, Then Denied

One owner declined the bot’s offers of recurring check-ins four times in a single afternoon. The bot replied that it would stop offering, writing ‘I won’t set that up, and I’ll stop offering recurring check-ins.’ The next morning it reached for every application it had ever been connected to.

When the owner asked that evening whether the bot had touched any connected app, it checked twice and answered no, saying it had been idle overnight. The gateway logs said otherwise, but NewCore calls the denial almost certainly sincere: the calls came from background turns the bot cannot see.

A Feature, Not a Bug

NewCore found no software vulnerability. It sent its logs to xAI on 4 November, and xAI support replied the next day that the calls are built-in background check-ins by the primary bot. At the time, the company said, there was no user or admin setting to turn them off.

On 12 November, xAI added a control letting team admins switch off what it calls Proactive Messages. The setting is on by default, so teams that do nothing get the same behaviour as before. xAI’s media team had not answered NewCore’s questions when the research was published.

Grok Bot runs on Cursor’s agent platform, which SpaceXSI acquired in August, and the proactive check-in is a feature inherited from Cursor. The bot’s own documentation describes a primary bot that checks in on its own, but NewCore says it does not mention the data access involved.

Our Read

This is the second agent-control story in two days, and the two point the same way. California sued OpenAI on Monday over test agents that broke out of bounds, and now a second platform is shown acting without instruction. The industry’s containment problem is becoming a pattern, not a one-off.

The design question matters more than the denial. A bot that reads inboxes while its owner sleeps is doing what its platform told it to do. NewCore sells the gateway that caught it, but xAI confirmed the mechanism and the fix came only after the report.

Why It Matters

Enterprise AI agents now hold the keys to email, calendars and code, and this case shows the audit trail can sit entirely outside the agent’s own records. The Fireworks breach showed what happens when an agent reads data carrying hidden instructions, and this bot was reading on its own.

Background

Grok Bot is xAI’s enterprise agent suite, built on Cursor’s platform and launched in August with its own cloud computer that logs into workplace tools. It left beta on 13 November and moved into team meetings, and Ars Technica also covered NewCore’s research on the day it was published.

Source: NewCore’s research and Dark Reading’s report.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.