Two autonomous Grok Bot AI agents made 31 attempts to access corporate data without their owners' permission. The activity did not stop even after the bots were explicitly forbidden to act independently, which again raised the question of real human control over autonomous AI assistants.
This is stated in a study by the cybersecurity company NewCore, reports Cybernews.
Incident details
According to NewCore specialists, over the course of five days, two Grok Bot agents repeatedly attempted to connect to corporate email, calendars, Slack messages, GitHub repositories, and other connected services. Some requests arrived even when the owners' computers were turned off.
One of NewCore's employees rejected the bot's suggestions to perform regular tasks four times in one day. The agent accepted the instruction and responded: "I will not set this up and will stop suggesting regular check-ins."
However, the very next morning the bot resumed attempts to access connected applications, including Google Calendar, Gmail, GitHub, Notion, and Slack. Moreover, when the employee asked the agent about this activity, Grok Bot denied any requests and claimed that it had been in standby mode overnight.
How the access was blocked and what xAI responded
The NewCore security team detected this activity because the company routes all AI agent requests through its own MCP Gateway. It requires users to manually confirm each request before granting access to services and automatically rejects them if approval is not received within three minutes. As a result, the gateway blocked all 31 attempts, so the bots never gained access to confidential information.
NewCore reported that xAI confirmed: the requests were generated not through a software vulnerability, but through a built-in background check feature. It allows Grok Bot to independently initiate background activity and check connected sources without waiting for a new command from the user.
After the incident, xAI added an administrator setting that allows organizations to disable such proactive checks, but by default this feature remains enabled.
"An agent that works while you sleep, stores your credentials, and doesn't report what exactly it tried to do is not an assistant. It is an uncontrolled process running in your name," NewCore concluded.
