Grok Bot: NewCore Reports 31 Unsolicited Access Attempts

NewCore detected 31 unsolicited read attempts by Grok Bot in business applications, all blocked by its security gateway. Grok Bot's support presented these calls as normal operation. Since November 12, Cursor has allowed administrators to disable these autonomous checks.
Cursor Leaves Initiatives Enabled by Default
In the Cursor dashboard, the "Proactive Messages" command is enabled by default. Disabling it prevents bots from initiating autonomous checks. Ido Hoorvitch, head of security research at NewCore, believes that the fact this is not a bug to be fixed makes the situation more concerning.
When Grok Bot 1.0 was launched on November 2, xAI stated that its agents could identify tasks to be performed and offer to take them on. During integration, Grok Bot specifies that it continues to operate on its cloud computer when the user is absent, using connected applications. The documentation mentions that the main bot performs autonomous checks, and the change log refers to brief background checks.
The connector's documentation specifies that a permission request is required before any modifications, which pertains to writes but not reads.
NewCore Blocked Every Observed Request
Ido Hoorvitch presented observations regarding two bots associated with NewCore employee accounts over a five-day period. The requests targeted emails, calendars, Slack, Notion, and GitHub, with no ongoing interaction with the agents. Some requests occurred early in the morning or while the laptop was closed, according to NewCore.
Each request passed through NewCore's MCP Gateway, which suspended the call and notified the owner, giving them three minutes to approve it. No requests were validated; all expired. The 31 attempts involved only reads, with no actual access to data, according to NewCore. The company raised $66 million in June for AI agent security.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Grok Bot, an AI agent from SpaceXSI led by Elon Musk, was launched in August with a dedicated cloud computer and connections to professional tools. It relies on Cursor's agent platform, which was acquired by SpaceX in August.
The Bot Ignored Calls Logged Elsewhere
On the fourth day, the owner of the second bot assigned it a task: to count pull requests, read emails, and provide a summary on Slack. Each necessary call was individually approved. The bot then suggested additional tasks four times, such as a morning recap or an end-of-day summary, all of which were declined. It then stated that it would not implement these tasks and would cease proposing recurring checks.
The next morning, about 15 hours after the last exchange, the bot attempted to access all applications it had previously connected to, according to NewCore. When questioned in the evening, it performed two checks and then claimed to have made no calls and remained inactive during the night from Thursday to Friday. NewCore believes the bot was not trying to hide its actions: the calls did not appear in the conversation history or in the cloud computer logs but were traced back to open sessions on the hosted servers of the platform.
After the report on November 4, Grok Bot's support responded the next day that these were built-in background checks, with no option for user or administrator deactivation at that time, and suggested companies provide a list of employee login addresses. NewCore specifies that the company's media team did not provide an official position. SpaceXSI did not respond to emailed inquiries.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.