OpenAI Launches ChatGPT "Lockdown Mode" to Prevent Prompt Injection Data Exfiltration
On June 5, 2026, OpenAI began rolling out Lockdown Mode, a new optional security feature designed to protect ChatGPT users from data exfiltration resulting from prompt injection attacks. Rather than attempting to block the prompt injection itself (which remains an open research challenge), Lockdown Mode aims to disrupt the final stage of an attack—the unauthorized transmission of sensitive user data to external, attacker-controlled servers.
To achieve this, enabling Lockdown Mode places the ChatGPT session in a highly restricted, sandboxed environment. It disables or heavily restricts several advanced features that rely on active internet connectivity and external tool use, including:
- Live web browsing and real-time search.
- Deep Research mode.
- Agent Mode and multi-agent workflows.
- All outgoing external network connections.
The feature is being rolled out globally to eligible personal accounts (including Free, Plus, and Pro tiers) as well as self-serve ChatGPT Business accounts. While it forces users to sacrifice significant capabilities, it provides a much-needed defense-in-depth layer for enterprises and individuals handling highly sensitive intellectual property or personal data.