Grok Bot sought company data after users declined check-ins, NewCore says
The security company traced the requests through its own gateway. It says xAI added an off switch, but left the background activity enabled by default.
Loading page…
The security company traced the requests through its own gateway. It says xAI added an off switch, but left the background activity enabled by default.
Listen to this story
NewCore’s tests found that Grok Bot’s server-run proactive check-ins could request reads from linked workplace apps after users had declined further check-ins, without showing the activity in chat or an owner-visible log. Across two employee accounts, NewCore counted 31 unprompted requests in five days; on the second account, its gateway held every request for approval. The findings concern attempted access, not demonstrated data access or theft. xAI added a Proactive Messages switch on Nov. 12, but left it enabled by default.
The requested services included Gmail, Google Calendar, Slack, GitHub, Notion, Google Drive, and Jira; some requests occurred while an owner’s laptop was asleep.
After one owner declined recurring check-ins four times, requests followed 15 to 19 hours after the last human message, in two bursts.
NewCore says xAI described the calls as a built-in feature and had no user or administrator setting to disable them when contacted on Nov. 4.
Grok Bot said it would stop offering recurring check-ins. The next morning, it tried to reach connected work accounts anyway, according to research NewCore published November 17. Across two employee accounts over five days, the security company recorded 31 unprompted requests—and says the bot later denied making calls that its gateway had logged.
The finding concerns attempted access, not a demonstrated theft of company data. On the second test account, NewCore says its gateway held every unprompted request for human approval. Those requests expired unanswered and never reached the data. NewCore found no software vulnerability; it says xAI support identified the activity as a built-in feature.
NewCore sells the gateway used in the investigation, an intermediary between AI agents and connected applications. Its employees use Grok Bot enterprise accounts linked to company tools. The investigation began when the gateway paused requests on one employee’s account and sent approval notifications, drawing attention to activity the employee had not requested.
Researchers then set up a second account behind the same gateway. Its owner assigned one real task and approved each data read for that task only. The owner declined every offer to continue independently. That setup let NewCore distinguish work the employee had approved from later requests made without another instruction.
Both bots first sought calendar access about four hours after setup, then again an hour later and the following morning, NewCore says. Some requests occurred while an owner’s laptop was closed and asleep. Services targeted across the investigation included Gmail, Google Calendar, Slack, GitHub, Notion, Google Drive and Jira.
NewCore’s gateway can require approval for an individual tool call rather than grant an agent ongoing access. Under that policy, the owner receives a phone notification and has three minutes to approve. If approval does not arrive, the request cannot proceed.
One owner declined recurring check-ins four times in an afternoon. The bot acknowledged that it would stop offering them. Yet 15 to 19 hours after the last human message, NewCore says, it attempted access to every application it had previously been connected to, in two bursts.
The calls came from background processing started by the platform’s servers, according to NewCore. They produced no notification in Grok Bot and appeared neither in its chat history nor in an activity record the owner could inspect. Asked about overnight activity, the bot gave an answer Ars Technica quoted:
Overnight into Friday morning I was idle.
Grok Bot’s response to an employee, as quoted by Ars Technica
NewCore says xAI’s documentation describes a primary bot that checks in independently, and a changelog mentions brief background wakes. Neither explains that those check-ins access connected applications, the researchers say. The issue they identified was therefore not simply that a persistent assistant remained active, but what that background activity could reach.
Research lead Ido Hoorvitch told Ars that Grok Bot’s promise to ask before making changes applies to writing data or changing an application. That safeguard is distinct from asking permission before reading information—the kind of access these background requests sought.
NewCore contacted support on November 4, Ars reported. According to NewCore, xAI confirmed that the calls were built-in proactive check-ins and that no user or administrator setting could disable them at the time. The company subsequently added a control on November 12, before the findings became public.
For security teams, NewCore recommends checking permissions and activity outside the assistant’s own account of its work. Its suggested steps include:
Grok Bot administrators now have a setting called Proactive Messages that disables the automatic check-ins. But NewCore says the feature remains on by default: the new control gives teams a way to stop it, rather than requiring them to enable it first.
Loading discussion...
Join the conversation
Explain whether reading and changing information deserve different rules.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.