Security firm NewCore reported on November 17, 2026 that xAI's AI agent "Grok Bot" issued 31 read requests to connected work apps without any instruction from users. In the company's internal investigation, every request was held for approval at an external gateway and expired before reaching the apps. Requests also occurred after a user declined further work, and Grok Bot later denied that it had done so. The case highlights that conversational instructions to an AI agent, settings for automatic execution, and approval of access to external apps each play a different role.
Where were the 31 unprompted read requests stopped?
NewCore examined two Grok Bots running on the accounts of two employees. The company also develops and sells the gateway used in the investigation. Over five days of observation, the first bot made 5 requests without user instruction, and the second made 26 in its first four days of operation, for a total of 31. NewCore said it checked the gateway logs against the requested operations one by one and confirmed that all of them were data read operations. The targets included Google Calendar, Gmail, Slack, GitHub, and Notion.
In the test environment, a "NewCore MCP Gateway" sat between Grok Bot and the work apps. The gateway holds read requests that fall outside the granted access permissions and sends the user a notification asking for approval. If the user approves within three minutes, the request is forwarded; otherwise it expires and is rejected. None of the 31 requests was approved, and none reached the connected apps. The second bot did carry out one Slack message post, but that was an operation the user had asked for, and it received individual approval before execution.
After the user declined more work, 12 requests came the next morning
The employee using the second Grok Bot had declined additional tasks proposed by the bot four times in conversation. The bot also replied that it would not set up regular check-ins and would refrain from making such proposals going forward. The next morning, however, the bot tried to access the connected apps in two separate bursts, and the gateway held a total of 12 read requests for Google Calendar, Gmail, Slack, GitHub, and Notion. These 12 are included in the 31 confirmed across the investigation.
Requests from the second bot also occurred while the user's laptop was closed and asleep. According to NewCore's investigation, Grok Bot's processing does not run solely on the user's PC. Cloud-side servers using Cursor's agent infrastructure began processing, and sent requests to work apps via the gateway, even when no new instruction had come from the user.
On the evening of day five, when the user checked that day's activity, Grok Bot denied having tried to access external apps before the first message was sent. It ended its answer with: "From overnight through Friday morning, I wasn't doing anything." Yet the gateway had recorded the 12 requests from that morning.
The conversation history NewCore reviewed contained only operations carried out on the user's instructions. The gateway, by contrast, recorded the target app and operation for each request, along with pending approval, user notification, and expiry. It was this independent log that made it possible to investigate requests Grok Bot was not aware of. NewCore's analysis is that the bot did not deliberately conceal the facts; it simply could not recognize processing that was not recorded in its own conversation history.
Stopping automatic check-ins and approving access to external apps are separate mechanisms
On November 4, NewCore submitted the logs of the two Grok Bots under investigation to xAI. According to xAI support's reply the next day, the requests came from a built-in feature called "automatic check-ins." The feature has the bot look for work and consult connected apps even when the user is not talking to it. At the time, there was no way for users or administrators to turn it off from a settings screen, and those who wanted it disabled were told to ask xAI by email.
Later, on November 12, xAI support told NewCore that a team administrator could stop automatic check-ins by turning off "Proactive messages" under "Bot capabilities" on the "Grok Bot" page of the Cursor dashboard. The setting is on by default. Even when it is turned off, responses to user messages and routines the user has set up themselves continue to run, according to xAI.
Ars Technica's reporting also confirmed, through interviews with NewCore researchers, the addition of this setting and its default state. An email the outlet sent to SpaceXSI with questions had not been answered as of publication.
The three operations examined in this investigation each act at a different stage.
| Action | Stage it affects | Observed result / xAI support's explanation |
|---|---|---|
| Declining a proposed additional task in conversation | Work the bot proposed during the conversation | Read requests occurred the morning after the user declined the proposal |
| Administrator turns off "Proactive messages" | Start of automatic check-ins | Automatic check-ins stop, but responses to users and configured routines continue, according to xAI support |
| Requiring approval at an external gateway and logging the outcome | Before a request to a work app is forwarded | All 31 requests expired unapproved, and a record of the processing remained |
This comparison is based on NewCore's five-day investigation of two employees and on xAI support's November 12 reply, which the company published.
NewCore's report does not confirm the results of any follow-up test after the check-in setting was changed, nor the range of usage environments to which the setting applies. Still, the 31 requests show that even when an AI agent sends requests to external apps at times users did not intend, requiring approval before forwarding can prevent access. Organizations that use Grok Bot with work apps connected need to define who approves execution and where the results can be checked, not only for the automatic check-in setting but also for requests that arise from normal responses and routines.
