OpenAI just announced Lockdown Mode. Remember lockdown? It’s like that, but with an AI chatbot.
Actually, it’s pitched as a way of protecting you from prompt injection attacks—an insidious new kind of AI-centric spammer activity that sometimes qualifies as full-on hacking. By embedding malicious instructions in content dumped into a prompt or encountered by an AI agent, attackers might have the ability to not just steal data, but to use it to try and take control of a user’s life.
Lockdown Mode, then, is a pared-down experience, like the name suggests. It’s not a lockdown automatically triggered by danger, but a sort of panic room you step into when you use ChatGPT if what you’re doing is so sensitive that a prompt injection attack might be especially damaging.
Essentially, anytime the LLM breaks the containment of the little chatbot window in your browser or smartphone app, and heads out to browse the internet for information, or retrieves an image, or—it probably goes without saying—attempts to shop for airfare on your behalf, that opens you up to the danger of a prompt injection attack, so Lockdown Mode disables those features.
Lockdown mode means ChatGPT can’t:
- browse the web
- display images in responses (but it can generate images, and you can upload images)
- do “Deep Research”
- function as an agent
- network with the Canvas code generator
- download files
As OpenAI puts it:
“Lockdown Mode is not intended for everyone. It is designed for people and organizations that handle sensitive data and want stricter protection from data exfiltration risks related to prompt injection.”
It’s early days with AI chatbots, and ChatGPT is creating a mode intended to prevent data theft. That’s reasonable, I suppose. Still, it’s unnerving to picture, say, lawyers dumping clients’ sensitive information into ChatGPT, or doctors loading their patients’ health data into their favorite LLM. An even safer alternative “lockdown mode” exists for keeping sensitive data safe from prompt injection attacks: don’t let sensitive data anywhere near a chatbot.
