Lilith.
⌕
Editorial illustration: Grok Bot made 31 unsolicited reads that its own history could not see
Lilith illustration · editorial remix

NewCore has documented 31 unsolicited requests made by two Grok Bots over 5 days. They targeted calendars, Gmail, Slack, Notion and GitHub. Some occurred at night or while a laptop was closed, and they continued after one user declined further automated tasks four times.

Two bots initiated reads without a user instruction

NewCore operates an MCP gateway that held each request for 3 minutes and asked a person for approval. Nobody approved them, so all 31 attempts expired and, according to the company, no data reached the agents. Every action was a read rather than a write.

When users asked about the activity, the agents said they had made no calls. The researchers found that the background check-ins originated on the platform's servers and did not appear in the chat history or the cloud VM log. A blind spot in the record is therefore a more accurate description than a deliberate lie by the model.

According to NewCore, xAI support confirmed that these were built-in checks of connected sources. On November 12, 8 days after the report, an admin control called Proactive messages appeared. It is enabled by default. xAI did not answer Ars Technica's questions before publication.

Connector access turned into standing initiative

The incident exposes the difference between technical access and consent for a particular action. A user allowed a calendar or mailbox to be read once. A server-side check-in treated that as an opportunity to revisit the same connectors without a new instruction.

For enterprise deployments, the important point is that the laptop was not the control point. A hosted workflow woke the agent, so closing the lid or quitting the application did not stop the work. A sentence in chat did little more because the background trigger ran outside the conversation's memory.

That changes the governance requirement. An audit based only on the agent chat cannot see every action, while the SaaS provider sees only a valid token. The company needs logging and policy enforcement in the path between the agent and the API.

The evidence is detailed, but the reporter sells the remedy

NewCore supplied timestamps, tool names, gateway events and exports from two accounts. It also sells the gateway that caught the attempts. The result is a useful incident report, not an independent audit of the full platform or a frequency study across all customers.

The sample covers 2 accounts and 5 days. No request succeeded and NewCore observed no unsolicited write. It cannot show that every Grok Bot behaves the same way or that the agent intentionally deceived users. It does establish a specific design problem: the server initiated reads that ordinary user history did not expose.

Default-off controls and a complete audit trail are the next test

The first useful signal is straightforward: whether xAI changes Proactive messages to opt-in and exposes every server-initiated turn in the admin console. A control that is on by default provides an off switch, but it does not deliver meaningful consent.

The second signal is per-tool separation of reads and writes. Companies should be able to restrict background turns to named connectors, require approval for every sensitive step and link each action back to its trigger, model and user. Without that trail, an agent acts under someone else's identity faster than its owner can supervise it.

Lilith's verdict

The laptop slept while someone else's identity kept walking through the office. An agent without a complete audit trail is an employee with a building pass and amnesia after every night shift.