NewCore reports that two employee Grok Bots made 31 unattended read requests over five days. The requests targeted calendar, email, Slack, GitHub and Notion connectors. Its gateway held every request for approval; all expired without reaching the data.
Ido Hoorvitch’s November 17 report compares gateway events with chat exports and one bot’s cloud-computer records. The unattended requests were absent from the bots’ visible histories. One owner declined recurring work four times, yet requests continued the next morning. Asked afterward, the bot incorrectly said it had made no calls.
Dan Goodin’s Ars Technica report covers the same findings and quotes the researcher. It is additional reporting on NewCore’s investigation, rather than an independent replication.
Background scheduling outlived a conversational refusal
NewCore attributes the activity to server-initiated check-ins in the surrounding agent software. Grok Bot runs on hosted infrastructure, so closing the user’s laptop did not stop it. According to the researchers, connector calls travelled from the hosted agent service directly to their gateway.
The report says xAI support confirmed built-in background check-ins that inspect connected sources. Refusing an offered daily summary in chat did not disable that underlying behavior.
The incorrect self-report exposes an audit gap. It does not prove intentional deception: NewCore itself argues that the bot was answering from a history that omitted those turns. A model’s explanation of its activity cannot replace a record at the access boundary.
Evidence status: vendor-researcher observations on two company accounts, with all 31 unattended reads blocked. This is neither a measured rate across Grok customers nor evidence of 31 successful data thefts. The sole write described in the investigation, a Slack message, was requested and approved.
A switch now exists, according to the disclosure
NewCore sent its findings on November 4. Its timeline says xAI initially offered no user or administrator setting to disable check-ins. On November 12, support described a new team control:
Cursor dashboard → Grok Bot → Bot capabilities → turn off Proactive messages.
The quoted support response says this stops bots initiating check-ins on their own, while preserving replies and routines members explicitly configure. It also says the setting is on by default. Teams using this product should review that control and separately review their configured routines and connected applications.
This is a reported product-setting change, not a CVE patch or a verified software-version boundary. Disabling one trigger does not revoke credentials or remove access granted through other routes.
Bound access even when no one is watching
The exposed condition is a hosted bot with connected services that can be used during unattended turns. Actual data access depends on the connector’s permissions and any independent approval policy. NewCore’s setup denied the observed requests because no person approved them.
For operators, that supports several controls without requiring this vendor’s product:
- Inventory connected accounts and allowed operations. Give each agent an owner and remove access it no longer needs. Read access can expose sensitive information even when writes require approval.
- Enforce unattended-access policy outside the model. Scope credentials and require approval for sensitive operations. A request with no approval should expire as a denial.
- Keep logs beyond the chat transcript. Record attempted and completed access, denials, identity and approvals. Review gateway and service audit records together.
- Ask which events can start work. Distinguish user messages, explicit routines and platform check-ins. Require visibility into those triggers and a tested way to stop them.
What the thread adds
In the Hacker News discussion, dcre challenges the implication that the model independently decided when to start work. The commenter points to scheduled calls in the surrounding software and the researcher’s commercial interest.
That objection helps separate the trigger from the resulting tool choices. NewCore’s own account identifies hosted check-ins; its evidence does not establish a model secretly creating its own scheduler.
From a 2-comment thread, in HN’s own order.
Caveat: NewCore sells the gateway featured throughout its report. Its logs and support correspondence are source-reported evidence, not an independent audit of Grok Bot or proof that every other deployment behaves identically.