Particle.news

OpenAI Rolls Out Lockdown Mode to Limit ChatGPT’s Web‑Connected Tools

The optional setting targets users who handle sensitive data by blocking outbound network requests to reduce the final‑stage risk of prompt‑injection data exfiltration.

Overview

  • OpenAI began rolling out Lockdown Mode in early June to eligible personal accounts and self‑serve ChatGPT Business subscriptions so users and workspace administrators can enable it from Settings.
  • When enabled, Lockdown Mode disables or limits features that make outbound network requests, including live web browsing (to cached content only), Deep Research, Agent Mode, image retrieval/display, Canvas networking, and file downloads.
  • OpenAI says the feature is meant as a last‑stage mitigation that lowers the chance an attacker can move stolen data off a session but does not stop prompt injections from appearing in content or guarantee complete prevention of exfiltration.
  • The rollout includes account controls such as an Active Sessions manager and works with workspace admin controls so administrators can restrict app and connector actions for members using Lockdown Mode.
  • Security researchers call the move a pragmatic but blunt tradeoff between safety and functionality and advise adding pre‑model scanners, secret scrubbing and tool‑result inspection to catch exfiltration attempts without disabling features entirely.