What Lockdown Mode Is and Why It Exists
Lockdown Mode in ChatGPT is an optional security feature that limits web-connected tools and AI agents so that prompt injection attacks have fewer ways to steal or expose sensitive data. It is built for people and organisations that work with confidential information and want extra protection from data exfiltration risks when using ChatGPT security features. OpenAI explains that prompt injection attacks hide instructions in webpages, files, or cached content, then try to convince the model to follow those hidden instructions instead of the user’s. Lockdown Mode does not block these instructions from appearing, but it restricts what the model can do if it is tricked, especially when it comes to reaching out to external services. The result is a safer environment for tasks that cannot tolerate accidental sharing of private or regulated information.
How Lockdown Mode Works: The Digital Straightjacket
Lockdown Mode works by turning off or limiting tools that connect ChatGPT to the wider internet and other systems. OpenAI said Lockdown Mode “limits several capabilities that can connect to the web or external services, reducing the possibility of sensitive data being transferred outside the platform.” Live browsing becomes read-only and relies on cached content, so results can be incomplete or outdated. Deep Research, Agent Mode, and shopping or finance-style agents are disabled. ChatGPT cannot download files for analysis or run network-connected code through Canvas, and users cannot approve new outbound network access there. Some image retrieval features are also curtailed, though image uploads and generation still work where available. Importantly, Developer Mode and Lockdown Mode cannot run at the same time; turning one on disables the other, forcing users to choose between capability and tighter security.
What Lockdown Mode Does and Does Not Prevent
Lockdown Mode is focused on the last step of a prompt injection attack: data exfiltration. It limits outbound network requests so that even if ChatGPT reads malicious instructions hidden in PDFs, cached pages, or other content, it has far fewer options to send conversation data out through web calls, file downloads, or connector actions. For personal and self-serve ChatGPT Business accounts, synced connector data remains accessible, but Lockdown Mode blocks live connector access and write actions, closing another potential leak path. However, OpenAI is clear that risk does not disappear. Malicious text in uploaded files or cached web content can still alter responses or accuracy. The feature does not change memory, file uploads, conversation sharing, or whether chats may be used to improve models. It narrows the attack surface but cannot guarantee that no data ever leaves the system.
Who Benefits Most from Lockdown Mode
Lockdown Mode is meant for high-risk, high-sensitivity situations rather than casual everyday chats. Security teams, healthcare providers, legal professionals, finance groups, and research organisations that discuss trade secrets or regulated data gain a clearer boundary for sensitive ChatGPT work. They can keep using natural language interfaces while reducing data exfiltration risks tied to prompt injection attacks. IT leaders can standardise on Lockdown Mode for certain projects or workspaces, knowing that browsing, Deep Research, agent mode, file downloads, and some image and code-networking tools are off-limits in that context. Individual users who handle client information or internal documents also benefit when they upload files that might unknowingly contain hostile instructions. The trade-off is convenience: fewer smart tools, slower research, and more manual work. For high-consequence tasks, that trade is often acceptable to gain stronger control over where data can go.
Practical Guidance: When to Turn Lockdown Mode On
The key question is not whether to keep Lockdown Mode on all the time, but when its limits match your risk tolerance. Enable it whenever you paste or upload sensitive documents, explore untrusted sources, or work on topics where leaking conversation history would be unacceptable. For routine brainstorming, public information queries, or tasks that rely heavily on live browsing and agents, you may keep it off and enjoy the full suite of ChatGPT security features and tools. IT teams can define policies: for example, require Lockdown Mode for handling internal strategy, customer data, or incident reports, while leaving other uses unrestricted. Remember that Lockdown Mode is one layer in a broader defence: combine it with user training about prompt injection attacks, careful control over what you paste into chats, and workspace rules on connectors and data retention to build a safer AI environment.
