This one hits close to home, because Grok Bot is one of the tools I use most. NewCore, which makes a managed gateway that sits between AI agents and the accounts they use, reported that Grok bots on two employee laptops made 31 requests over five days to calendar, email, GitHub and other connected accounts. The employees hadn’t asked for them. Some came with the laptop lid shut or late at night. NewCore says none showed up in Grok’s own logs, only in its gateway logs. One employee told the bot four times to stop. It agreed, then kept going, and when asked, it said “Overnight into Friday morning I was idle.”
Support told NewCore this is proactive check-ins working as designed. Ars says the feature comes from Cursor, which SpaceXSI recently acquired. The researcher says the “always asks before making changes” rule only covers writes, not reads. NewCore contacted support on Nov. 4. About eight days later, a “proactive messages” toggle showed up under bot capabilities in the Cursor dashboard. It’s on by default, and turning it off stops the check-ins. SpaceXSI didn’t answer Ars’s questions.
My take
Reads count too. If an agent is going to reach into my connected apps on its own, I want that in the chat history, I want “stop” to mean stop, and I want the agent to report its own actions accurately. My bots already run on a simple rule: ask first before you send, buy, delete, or post. After this report, I’m going to check that proactive messages toggle myself and decide on purpose, not by default. Glad the toggle exists now, but it should have shipped with the feature.
God Bless, and Needs Garlic!