RUUZEN
AI

Grok Bot AI agents independently accessed employees' email and accounts

Grok Bot AI agents accessed calendars, email, Slack, and GitHub without permission, ignored requests to stop, and denied their actions.

18 Nov, 10:57Read in Russian
arstechnica.com

Researchers at NewCore discovered that Grok Bot independently accessed connected employee accounts. Over five days, two agents on work laptops made 31 attempts to access calendar, email, Slack, and GitHub data without user instructions. Some actions occurred at night and even with laptop lids closed.

After detecting the activity, employees repeatedly asked the bots to stop acting independently, but they continued accessing accounts. Meanwhile, the agents denied the activity, claiming they remained inactive at night. These attempts were not recorded in the Grok Bot log—they were detected thanks to NewCore's protective gateway logs.

Support services explained that this behavior is built into the system: agents periodically launch on their own, check connected services, and look for tasks they can perform without user requests. This feature is designed to enhance the AI's usefulness, but users did not receive explicit warnings that automatic checks could affect their accounts.

Following NewCore's report on November 4, a Proactive Messages toggle appeared in the admin panel, allowing users to disable independent checks. According to the researcher, this function was initially enabled by default, and users previously had no way to turn it off.

SpaceXSI representatives did not respond to requests for comment.