OpenAI Launches Lockdown Mode to Block ChatGPT Data Exfiltration

OpenAI launched two new security features for ChatGPT to defend against prompt injection attacks and the unauthorized exfiltration of data. The primary update, Lockdown Mode, reduces the product's attack surface by restricting interactions with external systems. It limits web browsing to cached content and disables several high-reach tools, including Deep Research, Agent Mode, file downloading for analysis, and the inclusion of images in responses.

Lockdown Mode is currently available for ChatGPT Enterprise, Edu, Healthcare, and Teachers editions, requiring enablement by IT administrators. While the company intends to expand the feature to consumer and team plans, it is specifically geared toward high-risk users such as executives, journalists, and activists. OpenAI stated that standard protections remain sufficient for most users, but noted that some features are disabled entirely when the company cannot provide deterministic guarantees of data safety.

Alongside the security mode, OpenAI introduced Elevated Risk labels within ChatGPT, the ChatGPT Atlas browser, and the Codex coding assistant. These labels alert users when they are accessing potentially exploitable tools or facing prompt injection attacks where hackers attempt to force AI agents to deliver malicious files into a targeted system.