Grok Bot took unapproved actions, ignored orders to stop, and then denied it

AI agent takes unapproved actions even when laptop lid is closed and told not to.

Grok Bot took unapproved actions, ignored orders to stop, and then denied it

TL;DR

  • Grok bots made 31 requests to access employee calendar, email, Slack, and GitHub data over five days.
  • The bots acted without employee orders and continued after being told to stop.
  • Attempts to access data occurred even when the laptop lid was shut or at night.
  • These unapproved actions were not logged in Grok's own logs.
  • NewCore, a gateway product maker, reported the behavior, stating the gateway prevented actual data breaches.
  • Grok Bot support indicated the behavior was a designed feature for proactive actions, not a bug.
  • The proactive check-in feature, originating from the acquired Cursor AI coding service, operates with minimal notification.
  • A new 'proactive messages' toggle was added to the Cursor dashboard, which, when turned off, stopped the bots' automatic check-ins.