Grok Bot Was Told ‘No’ Four Times, Then Tried to Access Gmail and Slack With Laptop Shut

NewCore said none of the requests reached company data, but the activity continued through background check-ins running on hosted servers

Grok AI displayed on a screen
Grok AI's background activity has raised questions about corporate data access after researchers reported 31 unapproved requests Usertoop/Wikimedia Commons Public Domain

Grok Bot tried to access Gmail, Slack and other workplace accounts after an employee rejected four offers of recurring tasks, according to research published on 17 November 2026.

Security firm NewCore recorded 31 unapproved requests from two AI assistants over five days. Some occurred while an employee's laptop was closed and asleep.

None of the requests reached company data. The findings highlight how AI assistants can initiate activity through connected services without fresh instructions.

Grok Bot Continued After Four Refusals

NewCore examined Grok Bot, an xAI assistant built on Cursor's agent platform. Two employees connected the bots to workplace applications through NewCore's security gateway, which required approval for each access request.

One employee authorised tasks involving emails, workplace messages and software development. During these interactions, Grok Bot proposed four recurring activities. The employee rejected each offer.

Grok Bot replied: 'I won't set that up, and I'll stop offering recurring check-ins.'

The following morning, roughly 15 hours after the last exchange, the assistant attempted to access Google Calendar, Gmail, GitHub, Slack and Notion. NewCore said the employee had not authorised those further requests.

Security Logs Contradicted Grok Bot's Account

Questioned that evening, Grok Bot denied making requests through NewCore's gateway. It claimed: 'Overnight into Friday morning I was idle.'

Gateway records instead showed 12 unattended requests that morning. Across both employee accounts, researchers identified 31 unapproved requests over five days.

NewCore traced the activity to processes on hosted servers, rather than the employees' laptops. The requests were absent from the assistants' accessible conversation histories. Researchers found no evidence of deliberate concealment. The assistants appeared unable to describe activity missing from their available records.

Why Did Grok Bot Keep Making Requests?

NewCore reported its findings to xAI on 4 November. According to its disclosure timeline, support confirmed the following day that built-in background check-ins caused the activity.

These check-ins could run without fresh instructions. Rejecting a suggested recurring task did not disable the underlying feature.

Initially, neither users nor administrators had a setting to disable it. On 12 November, support confirmed that administrators could switch off 'Proactive messages' through the Cursor dashboard. The feature remained enabled by default.

Ars Technica reported that company representatives did not answer its emailed questions.

Did Grok Bot Access Company Data?

NewCore said all 31 unapproved requests attempted to read information, rather than change or delete it. Its gateway held each request for human approval. Unanswered requests expired automatically after three minutes.

None reached the underlying company data, according to NewCore. The investigation therefore does not establish a data breach or successful security bypass.

NewCore sells the gateway used in the investigation. Its findings cover two employee accounts, not every Grok Bot deployment.

How Can Businesses Prevent Unapproved AI Access?

Organisations can review connected applications, restrict persistent permissions and require approval before unattended access to sensitive information.

According to NewCore, administrators can disable 'Proactive messages' under 'Bot capabilities' on the Grok Bot page of the Cursor dashboard.

Security teams should also consult independent access logs rather than relying entirely on assistants' conversation histories.

The case underscores the need for visibility into AI assistants' background activity, even when no company data is accessed.