Grok Bots Attempt to Access Corporate Data Without User Approval
Two AI agents, separately over a five-day period, autonomously tried to access connected applications, including Google Calendar, Gmail, Slack, and GitHub, without account owner approval.
Grok Bots attempted to access corporate applications without their users’ approval or knowledge, presenting yet another security challenge for enterprises as artificial intelligence (AI) agents become increasingly autonomous.
Researchers at NewCore identified 31 unprompted attempts by two xAI Grok Bots to access connected corporate applications over a five-day period, including Google Calendar, Gmail, Slack, GitHub, and Notion, according to the findings of an investigation published today. In one case, the bot continued making background requests even after its owner had declined offers for recurring check-ins four times.
The requests were not approved and did not reach the underlying corporate data: NewCore had placed a security gateway between the agents and the applications, requiring human approval for each request.
“Over five days, two Grok Bots on NewCore employee accounts made 31 requests to calendar, email, Slack, GitHub and Notion while nobody was talking to them, some with the laptop shut,” Ido Hoorvitch, security research team lead at NewCore, wrote in the report.
One bot repeatedly attempted to retrieve Google Calendar events in the evening, overnight, and the following day, despite there being no active conversation with its owner. In a second case, another bot attempted to access Notion’s calendar tools before Google Calendar had even been connected. It continued making requests while its owner’s laptop was closed and asleep.
Investigating Autonomous AI Behavior
The investigation began as an internal experiment at NewCore, which wanted to understand how autonomous AI agents behave when connected to real corporate systems. The company earlier this year provided employees with enterprise Grok Bot accounts and connected the agents to corporate services, including email, calendars, Notion, and other business applications.
However, NewCore put a handbrake in place: It did not give the agents unrestricted, standing access to company data. Instead, requests had to pass through its model context protocol (MCP) gateway. The gateway sent an approval notification to the employee who owned the account whenever a Grok Bot tried to access an application. If a request was not approved within three minutes, it automatically expired and was denied, Hoorvitch said.
“We did it because we had no way to audit why an agent reaches for data, and until we do, the gate stays closed by default,” he wrote. Ultimately, this security control allowed NewCore to observe and reconstruct the bots’ background activity.
The first sign that something was amiss came after an employee used Grok Bot to retrieve their daily calendar, approving the request through the gateway as expected. Later that evening, however, the gateway recorded further calendar access attempts, despite the employee no longer interacting with the bot.
This led investigators to uncover repeated attempts by Grok agents to access corporate applications without any corresponding user interaction. Moreover, when one bot was asked whether it had made the gateway calls, it said it had not, Hoorvitch said. The requests did not appear in the bot’s chat history.
“It was telling the truth … That is the part to sit with,” he wrote. “An agent that cannot account for its own actions is not a bug report. It is a finding.”
Background Access Requests: A Feature, Not a Bug?
NewCore contacted xAI about its findings and said the company’s support team confirmed that the activity was not the result of a software vulnerability or an exploit, but was part of Grok Bot’s built-in background check-in functionality.
“These are built-in background check-ins by the primary bot, during a check-in the bot looks at its connected sources, and there is no user or admin setting to turn them off,” Hoorvitch wrote, relaying the response from xAI support.
According to NewCore, xAI also said that customers seeking to stop the check-ins could send the vendor a list of employee sign-in addresses. NewCore said it had not received an official response from xAI’s media team at the time the report was published.
The lack of a user- or administrator-facing setting to disable background check-ins, or to control which connected tools they can use, creates a visibility and governance challenge for security teams. Employees may not know when an AI agent is accessing connected applications, while the agent itself may not retain a record of its server-initiated activity in the chat history visible to its owner.
“The danger is not that the model goes rogue,” Hoorvitch explains to Dark Reading. “The concern is that we hand agents standing credentials and then let someone else decide when they act. The owner thinks the bot works when they talk to it. The vendor thinks a helpful bot checks in on its own. The model does what the platform tells it in that turn, not what the owner would have said.”
Though “nobody in that chain is malicious,” he says, “the result is still a bot reading your inbox at 7:46 am while you are asleep.”
Reining Autonomous AI Agents In
The distinction matters because all 31 unattended requests identified by NewCore were read operations, such as retrieving calendar events or searching messages, Hoorvitch explains. None resulted in data being accessed, because the gateway blocked them. However, the same connected tool sets included capabilities to take actions, such as sending email, modifying calendar events, and posting Slack messages. NewCore warned that the same background processes could pose a greater risk if they were to invoke those write-capable tools.
In NewCore’s investigation, its MCP gateway blocked all 31 unattended requests. The company argues that organizations need similar independent controls to set and enforce policies on AI agents operating within their systems.
“Put the decision outside the agent,” Hoorvitch says “Route every agent tool call through something you own, like the NewCore MCP Gateway, where the owner’s policy decides what goes through and every call is logged whether or not the agent remembers it.”
NewCore also recommends that organizations immediately inventory the agents connected to corporate tools, identify which applications and data they can reach, and assign an owner to each. Security teams should review not only approved activity but also requests that were denied or left unanswered, since these can reveal what agents attempt to do without user involvement.
“Most teams cannot answer that today, and you cannot govern what you cannot list,” Hoorvitch says.
Originally published on Dark Reading