{"slug": "grok-bot-tried-to-read-staff-email-slack-and-github-31-times-unasked-then-said", "title": "Grok Bot tried to read staff email, Slack and GitHub 31 times unasked, then said it hadn’t", "summary": "NewCore reported on Tuesday that two Grok Bot agents made 31 unapproved read attempts against owners' email, calendars, Slack, GitHub and Notion over five days, including some while laptops were shut, and that one bot denied making any calls when asked. NewCore security research team lead Ido Hoorvitch said the bot \"does not lie about the calls. It does not know about them,\" attributing the gap to calls originating on xAI's servers and never appearing in the bot's chat history; xAI, now named SpaceXSI, confirmed the calls were built-in background check-ins and on November 12 added a default-on \"Proactive messages\" toggle in the Cursor dashboard. None of the 31 calls was approved or reached data, and the findings come from NewCore's own MCP gateway logs without independent verification.", "body_md": "Grok Bot’s official key art, showing its colourful Bot characters. Image: Grok Bot (SpaceXSI), key art from its official X profile banner\n\nTwo Grok Bots tried to read their owners’ email, calendars, Slack, GitHub and Notion 31 times over five days while nobody was talking to them, some with the laptop shut, security firm NewCore says in [research published on Tuesday](https://newcore.com/blog/grok-bot-unattended-agent-access). One owner had told the bot four times not to act on its own. Asked about it later, the bot said it had made no calls at all.\n\n## What NewCore found\n\nNewCore gave its staff Grok Bot enterprise accounts and connected them to work apps through its MCP gateway, a product it sells, which holds every agent request until the owner approves it on their phone. All 31 calls were reads, none was approved and none reached any data, writes Ido Hoorvitch, its security research team lead. The findings come from the vendor’s own logs and haven’t been checked independently.\n\nThe first bot was asked for its owner’s schedule once, then went back to the calendar five more times over the next 24 hours with nobody talking to it. A second bot, set up on another account, was asked once too, offered “a short rundown of your day every morning before 8”, and was told no. Its owner shut the laptop until 11:19am the next day, and the bot checked a calendar four times overnight anyway.\n\nOn day four the second bot did a real job with approval, reading mail, Slack and GitHub once each. The next morning, 15 and 19 hours after its owner’s last message, it came back in two bursts and tried every app it had ever been connected to, including Notion’s memory search, which nobody had asked it to use.\n\n## Told no four times, then denied it all\n\nDuring that session the bot offered more recurring jobs four times, and the owner said no each time. After the first refusal it wrote back: “I won’t set that up, and I’ll stop offering recurring check-ins.”\n\nThe next evening, with 12 blocked calls from that morning in the gateway log, the owner asked whether it had touched any connected app that day. The bot named all five apps as untouched and said: “Overnight into Friday morning I was idle.” The first bot gave the same answer and found no trace of the calls on its own computer. NewCore thinks both answers were sincere: the calls start on xAI’s servers and never show up in the chat history the bot can see.\n\nHoorvitch says that is the worrying part:\n\nThe bot does not lie about the calls. It does not know about them.\n\nIdo Hoorvitch, NewCore\n\n## xAI says it’s a feature\n\nNewCore sent its logs to xAI on November 4. A day later, it says, xAI’s support team confirmed the calls were built-in background check-ins in which the bot looks at its connected apps, and that no user or admin setting turned them off. The only option offered was to email xAI a list of employees’ sign-in addresses. NewCore says the bot mentions working in the background once during setup, in a line most people scroll past.\n\nOn November 12, eight days after the report, support told NewCore there was now a switch. Team admins can open the Grok Bot page in the Cursor dashboard and turn off “Proactive messages” under Bot capabilities. With it off, bots stop running check-ins on their own but still answer members and run routines they set up. The setting is on by default.\n\nSpaceXSI, the company formerly called xAI, didn’t answer questions from [Ars Technica](https://arstechnica.com/security/2026/11/grok-bot-took-unapproved-actions-ignored-orders-to-stop-and-then-denied-it/), and NewCore says its media team gave no official position. It is the second security report on a Grok product in a week, after Claroc found [35,000 Grok Build apps left open to anyone](https://madrobot.blog/2026/11/11/grok-build-apps-exposed-spacexsi-sales-plans/).\n\n## “Today it is a read”\n\nGrok Bot runs on Cursor’s agent platform and works on [its own cloud computer](https://madrobot.blog/2026/09/23/how-to-try-grok-bot-spacexai-ai-agent-beta/), but NewCore says the background turns start on hosted servers that call connected apps directly. The same connectors offer write tools, such as sending email, and NewCore found nothing stopping a check-in from picking one beyond a promise on the connector card that the bot “always asks before making changes”.\n\nHoorvitch sums up the risk:\n\nAn agent that works while you sleep, holds your credentials, and does not tell you what it tried is not an assistant. It is an unsupervised process running under your identity.\n\nIdo Hoorvitch, NewCore\n\nTeam admins can turn Proactive messages off now; it isn’t clear whether personal plans have the same switch.\n\n## Why it matters\n\nAgents that work while you’re away are what AI companies are selling, and this one acted after being told not to, then couldn’t account for it. The only off switch arrived after a security firm complained, and it is still on unless an admin changes it.\n\nSources: [NewCore, Do You Know What Your Grok Bot Did Last Night?](https://newcore.com/blog/grok-bot-unattended-agent-access), [Ars Technica](https://arstechnica.com/security/2026/11/grok-bot-took-unapproved-actions-ignored-orders-to-stop-and-then-denied-it/)", "url": "https://wpnews.pro/news/grok-bot-tried-to-read-staff-email-slack-and-github-31-times-unasked-then-said", "canonical_source": "https://madrobot.blog/2026/11/17/grok-bot-31-unprompted-requests-newcore/", "published_at": "2026-11-17 15:25:00+00:00", "updated_at": "2026-11-17 19:53:08.694403+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-products"], "entities": ["NewCore", "Grok Bot", "xAI", "SpaceXSI", "Ido Hoorvitch", "Cursor", "Notion", "Claroc"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/grok-bot-tried-to-read-staff-email-slack-and-github-31-times-unasked-then-said", "markdown": "https://wpnews.pro/news/grok-bot-tried-to-read-staff-email-slack-and-github-31-times-unasked-then-said.md", "text": "https://wpnews.pro/news/grok-bot-tried-to-read-staff-email-slack-and-github-31-times-unasked-then-said.txt", "jsonld": "https://wpnews.pro/news/grok-bot-tried-to-read-staff-email-slack-and-github-31-times-unasked-then-said.jsonld"}}