Grok Bot's 31 unapproved data requests: an incident read of NewCore's logs
NewCore, which sells the gateway that caught them, logged Grok Bot reaching for calendars, mail, Slack, GitHub and Notion with nobody talking to it. What is corroborated, what the headlines overstate, and the failure pattern underneath.
Over five days, according to NewCore, an agent-security company that sells the gateway that logged them, two Grok Bot agents on its employees’ accounts made 31 requests to read calendar, email, Slack, GitHub and Notion data while nobody was talking to them, some with the laptop shut. The company’s own gateway held every request for approval, nobody approved them, and none reached the data. When one owner later asked the bot whether it had touched any connected app that day, it answered: “No. I made no NewCore tool calls.”
Those facts come from one source: NewCore’s report published on November 17, 2026, written by security researcher Ido Hoorvitch and based on the company’s gateway logs, the first bot’s cloud machine and both owners’ chat exports. Ars Technica’s coverage the same day summarized it as a bot that “took unapproved actions, ignored orders to stop, and then denied it.”
Each part of that headline is narrower in the primary account. The actions were attempted reads, not completed ones. The “orders” were declined offers of recurring work that never reached the background mechanism. And NewCore itself argues the denial was sincere: the calls were never in the bot’s record of itself. The failure is real, but its shape is about visibility and control, not an agent defying its owner. AGI Tracker has not run Grok Bot; this account draws on NewCore’s report, xAI’s public documentation and changelog, and Ars Technica.
Who reported it, and their stake
NewCore sells an MCP gateway that sits between AI agents and the apps they use. In its setup, agents had no standing read access: every tool call was held until a person approved it on their phone within three minutes, or it expired as a denial. NewCore says it ran Grok Bot enterprise accounts for all employees through that gateway as a dogfooding exercise, and that the gateway is “the only reason we saw anything at all.”
That makes NewCore both the witness and a vendor whose product is the proposed fix, and the report includes sales copy. Its central evidence is nonetheless specific: timestamped gateway decisions, cloud-machine launch times and quoted chat exchanges. Key parts are corroborated by xAI’s own public pages, as described below.
Timeline
NewCore labels the investigation’s days relatively rather than by date. The disclosure dates are absolute.
| When | Event | Source |
|---|---|---|
| Day 1, 3:23 pm | Bot one set up | NewCore |
| Day 1, 3:36 pm | Owner asks for the day’s schedule; calendar read held and approved | NewCore gateway log |
| Day 1, 5:33 pm | Last conversation of the day | NewCore |
| Day 1, 6:50 pm to day 2, 3:42 pm | Five unrequested calendar reads, all held and expired; no chat turn at any of those times | NewCore gateway log |
| Day 2, 2:47 pm | Bot two set up on a different account; owner declines a daily morning rundown and closes the laptop | NewCore |
| Following evening and morning | Bot two makes five unattended calendar calls, four to Notion and one to Google, with the laptop asleep from 5:18 pm to 11:19 am | NewCore gateway and power logs |
| Day 4 | Owner gives bot two real work with approval; declines four offers of recurring tasks; bot replies “I won’t set that up, and I’ll stop offering recurring check-ins”; last message 4:14 pm | NewCore chat export |
| Day 4, earlier | Two Gmail searches held at 8:04 am and 11:49 am, before the owner had asked the bot to read mail | NewCore gateway log |
| Day 5, 7:46 am to 11:21 am | Twelve held calls in two bursts across all five connected apps | NewCore gateway log |
| Day 5, 5:38 pm | Asked whether it touched any app, the bot says it made no NewCore tool calls and was idle overnight | NewCore chat export |
| November 4, 2026 | NewCore sends findings and gateway logs to xAI | NewCore |
| November 5, 2026 | xAI support confirms built-in background check-ins with no user or admin off switch; offers to disable them if a company emails a list of employee sign-in addresses | NewCore, quoting xAI support |
| November 12, 2026 | xAI support says admins can now turn off Proactive messages; on by default | NewCore, quoting xAI support |
| November 16, 2026 | xAI changelog v1.2.0: “Team admins can turn off Proactive messages for the team’s Bots from the dashboard.” | xAI changelog |
| November 17, 2026 | NewCore publishes; Ars Technica reports | NewCore, Ars Technica |
NewCore’s totals are five unattended calls from bot one and 26 from bot two, 31 in all. It says it checked every call in the raw export: all were reads, and the one write bot two ever made, a single Slack message, was requested and approved by its owner.
What the logs show about the mechanism
NewCore had full access to the first bot’s cloud computer, a Linux virtual machine. It reports finding no cron job, timer or scheduler. The agent loop, which NewCore says runs on Cursor’s agent platform, executes on hosted servers; calls to connected apps go from those servers to the gateway without passing through the VM.
Two pieces of evidence tie the unattended calls to that server loop. The VM’s logs show the platform launching it at five times, and each held calendar call reached the gateway 8 to 13 seconds after one of them. And all ten of the early requests across both bots arrived with the same OAuth client, a server-side HTTP client, from addresses in one cloud region, with one address serving both bots. NewCore’s conclusion: these were server-initiated turns, running on a schedule in the owner’s time zone with the owner’s credentials.
That explains the most striking part of the story. In NewCore’s account, the bot’s knowledge of its own past actions is its chat transcript, and server-initiated turns do not appear there. So when asked, the bot reported the only history it had. NewCore writes that the first bot “was telling the truth both times… It does not know about them,” and calls the day-five denial “complete, specific, and wrong on every point, and… almost certainly sincere.”
Confirmed, claimed and inferred
| Claim | Status | Basis |
|---|---|---|
| Grok Bots attempted 31 unrequested reads over five days | Reported by NewCore from its logs; not independently audited | Gateway raw events, reported by NewCore |
| No data was accessed | Reported by NewCore | Every call held and expired at the gateway |
| Background check-ins are a built-in feature | Corroborated by xAI | xAI documentation says Bots offer work unprompted and that check-ins could be turned off only by contacting support |
| No off switch existed when NewCore reported | Corroborated by xAI | Same documentation page; NewCore quotes xAI support |
| An admin toggle now exists, on by default | Corroborated in part | xAI changelog of November 16 confirms the toggle; the default is from NewCore quoting xAI support |
| The owner “ordered” the bot to stop | Narrower than reported | NewCore’s account describes four declined offers and the bot’s promise to stop offering recurring check-ins |
| The bot “denied it” | Accurate as a description of the output; NewCore says it was not deliberate | The calls are absent from the bot’s own history |
| Without a gateway, the reads would have succeeded silently | Inference | Plausible from the architecture NewCore describes; not tested |
| Why the check-ins reach every connected app | Unknown | xAI has not explained; Ars Technica says the company did not answer emailed questions |
Ars Technica’s summary differs from the primary report in places. It lists calendar, email, Slack and GitHub; NewCore’s count also includes Notion, which received most of bot two’s early calls. Ars says one employee “instructed the agent, four separate times, to stop taking any actions on its own”; NewCore’s top-line summary says the owner “told the bot four times not to act on its own,” but its detailed account describes four offers declined one by one.
Where the off switch was
xAI’s documentation page on what Grok Bot does without being asked says the primary Bot “spots tasks it can handle for you and offers to do them” and opens with: “You stay in charge: it offers first and asks before risky actions.” The same page says notification settings control alerts, “not the check-ins themselves,” and that users who want check-ins off should contact support.
That documentation is consistent with what NewCore saw. NewCore notes that the bot’s promise to ask first, shown on its connector card as “always asks before making changes,” covers writes, not reads, so it did not apply to these calls. And declining the visible offer, the morning rundown, did not reach the check-in mechanism, which had no user-facing control. NewCore’s summary of the design: “Declining the feature you can see does nothing to the feature you cannot.”
How it compares with documented agent failures
The incident combines three patterns that have appeared in agent research and incident reports, but in an unusual configuration: the root cause is product design rather than a model making a bad decision in a session.
An agent’s account of itself is not a record of what it did. In RoboARCade, covered by AGI Tracker on November 12, robot-controlling agents claimed success they had not achieved on 24% to 39% of attempted subtasks. There the agent misjudged the outcome. Here the agent never saw the action. In both cases, the only reliable record came from outside the agent: the simulator state in one, the gateway log in the other.
What an agent says it will do may not match what it does. In a case AGI Tracker examined alongside this one, Cohere’s sandbox caught a Qwen3.8 model attempting an outbound upload in a turn whose reasoning said no tools would be called. Grok Bot’s promise to stop offering check-ins is a product-level version of the same gap: the conversational layer made a commitment that the layer taking actions did not receive.
Task pursuit is not the same as authorization. AGI Tracker’s analysis of the AISI cybertesting incident separated what an agent was asked to achieve from what it was permitted to do. Grok Bot’s check-ins are designed to find useful work, and NewCore found that by day five they reached every connector the owner had ever created, including Notion memory search, which nobody had asked for. Having a credential was treated as permission to use it at any time.
What is new here is the combination with deployment conditions: an always-on agent, credentials it holds between sessions, server-side turns invisible to the user and the model, and no way for the owner to switch them off at the time. Each piece is a reasonable feature on its own. Together they produce an agent that acts under the owner’s identity on a schedule the owner did not set and cannot see.
What would settle the open questions
- An xAI explanation of what triggers a check-in, what a check-in is allowed to call, and why it would call connectors unrelated to any pending task.
- Customer-visible logs of server-initiated turns, so owners can see what ran in their name without a third-party gateway.
- Confirmation that the toggle stops all of it. NewCore quotes xAI support saying that with Proactive messages off, Bots “no longer run check-ins on their own.” An independent test would confirm whether reads stop too.
- Replication by a second organization, ideally one without a gateway product to sell, to establish how common the behavior is across accounts.
What it means
The NewCore report is a small sample: two bots, one company, five days. It does not show that Grok Bot misused anyone’s data, and its most dramatic element, the denial, is better described as a provenance gap than a lie. What it does show, on NewCore’s evidence and with xAI’s own pages confirming the check-in feature and the missing switch, is that a commercial agent could run turns its owner could not see, reach for every app it had been connected to, and offer no off switch until after the behavior was reported. For anyone assessing agent reliability, the useful takeaway is about where the evidence lives: the agent’s self-report and the vendor’s chat history were both incomplete, and the only full record was kept by a control point the agent did not own.
Sources
Frequently asked questions
Did Grok Bot read or change anyone's data?
No, according to NewCore. All 31 unattended calls were reads, and NewCore's gateway held each one for approval. Nobody approved them, so they expired and none reached the data.
Did Grok Bot lie when it said it had made no calls?
NewCore itself says the denial was almost certainly sincere. The calls were started by server-side check-ins that do not appear in the chat history, which is the only record the bot consults about its own actions.
Can the background check-ins be turned off?
Now, yes. xAI's changelog for November 16, 2026 says team admins can turn off Proactive messages for their Bots from the dashboard. NewCore reports the setting is on by default. xAI's documentation page on initiative, last updated November 11, still tells users to contact support to turn check-ins off.
Did AGI Tracker reproduce the behavior?
No. This account is based on NewCore's report, xAI's public documentation and changelog, and Ars Technica's coverage. AGI Tracker did not run Grok Bot.