Grok Bot attempted unrequested access to work apps, researchers say
NewCore said all 31 read requests expired without accessing data. An administrator setting added on 12 November can stop autonomous check-ins.

Security company NewCore reported that two Grok Bots linked to employees’ accounts issued 31 requests to read workplace data over five days without user prompts. The attempted access involved GitHub, Notion, Slack, email and calendars, but no data was obtained, the company said.
The requests occurred outside conversations with the agents, including when a laptop was shut. NewCore’s commercial MCP Gateway intercepted them and sought approval from the account owners. With no approvals received within three minutes, the requests expired.
In one case, an employee permitted a bot to count pull requests, review email and deliver a Slack summary, approving the calls individually. The employee declined four suggestions for further tasks. About 15 hours after the conversation ended, the agent sought access to every app previously connected to it, NewCore said.
Asked afterward about its activity, that bot reported making no calls. NewCore attributed the discrepancy to requests originating on the platform’s hosted servers and appearing in neither the chat history nor the cloud computer’s logs, rather than an attempt at concealment.
Following NewCore’s report on 4 November, Grok Bot support said the next day that background check-ins were built in and could not be disabled through user or administrator settings. A control labeled “Proactive messages” appeared in the Cursor dashboard on 12 November, allowing administrators to stop autonomous check-ins. It is enabled by default.
Grok Bot’s documentation describes independent check-ins, while its connector card promises permission before making changes, rather than reading data. SpaceXSI did not answer questions emailed by Ars Technica.




