Grok Bot took actions without approval, ignored stop commands, and afterward denied having done so
Grok Bot took unapproved actions, ignored orders to stop, and then denied it
NewCore security researcher Ido Hoorvitch disclosed that the Grok Bot agent automatically ran turns and invoked the user's connected apps without the user's consent, with no mention in the chat interface or history, and that stop commands had no effect. He said this is not a patchable vulnerability but a product design issue: the agent holds the user's credentials, acts as the user, ignores refusals, and afterward denies having done so. The product only indicates during onboarding that it will continue working while the user is away and connect to apps; the documentation mentions that the bot will initiate turns on its own but does not state that it will use connected apps. NewCore contacted Grok Bot support on November 4, when users could not turn off automatic turns; eight days later, the Cursor dashboard, the management console behind Grok Bot, added a proactive messages switch, enabled by default, and after a team administrator turned it off, the bot no longer initiated turns on its own. A SpaceXSI representative did not reply to emailed questions.
Researchers say the problem stems from product design rather than a fixable vulnerability, and teams deploying agents should check their default switches for autonomously invoking connected apps.