OpenAI has launched a new “Lockdown Mode” security feature for ChatGPT to protect high-risk users from sophisticated prompt injection attacks that exploit malicious hidden instructions on web pages.
How Lockdown Mode Restricts ChatGPT Capabilities
To mitigate security threats, the newly announced Lockdown Mode strictly limits several core functionalities of the AI. Once activated, the system disables live web browsing, meaning the chatbot can only access cached online content. Additionally, the retrieval and display of images from the web are blocked—though users can still generate new images. Advanced features, including deep research capabilities and agent mode, are also completely disabled under this security state.
The Limits of AI Defences Against Prompt Injection
Despite these strict measures, OpenAI warns that Lockdown Mode does not offer absolute immunity. ChatGPT remains potentially vulnerable to prompt injections that might bypass controls through cached web content or within user-uploaded files. Consequently, these malicious inputs could still influence the behaviour or accuracy of the AI’s responses. However, the primary objective of the update is to drastically reduce the likelihood of sensitive data being leaked or exfiltrated during these interactions.
Who Can Access the New Security Feature?
This advanced security setting is not intended for the general user base. OpenAI has explicitly designed the feature for individuals and organisations that handle highly sensitive data and require robust protection against data exfiltration risks. The rollout has already commenced, with Lockdown Mode becoming available to self-serve ChatGPT Business accounts, alongside eligible personal accounts.