Updated
Updated · TechCrunch · Jun 6
OpenAI Rolls Out ChatGPT Lockdown Mode, Disabling 4 Features to Curb Prompt Injection
Updated
Updated · TechCrunch · Jun 6

OpenAI Rolls Out ChatGPT Lockdown Mode, Disabling 4 Features to Curb Prompt Injection

3 articles · Updated · TechCrunch · Jun 6

Summary

  • OpenAI is rolling out Lockdown Mode for self-serve ChatGPT Business accounts and eligible personal accounts, targeting users that handle sensitive data and want tighter protection from prompt-injection attacks.
  • The mode disables 4 capabilities—live web browsing, web image retrieval and display, deep research, and agent mode—to reduce the chances that hidden instructions in webpages or other sources can trigger data exfiltration.
  • OpenAI said the safeguard is not foolproof: prompt injections can still appear in cached web content or uploaded files and may still affect a response’s behavior or accuracy.
  • The company framed Lockdown Mode as a stricter, opt-in setting rather than a default for all users, aimed at lowering risk for higher-sensitivity workflows.

Insights

Does ChatGPT's new security mode signal a safer AI, or a clever upsell to its premium Enterprise version?
With prompt injection an 'unsolved problem,' can any security mode truly prevent AI from being weaponized?