OpenAI is rolling out Lockdown Mode in ChatGPT to Business users and eligible personal accounts as an optional security setting for people at higher risk of prompt-injection attacks.

It disables live web browsing, web image retrieval in responses, deep research, agent mode, and other network-enabled features. It also limits ChatGPT to cached content so no live network requests leave OpenAI’s controlled environment.

The safeguard targets attacks in which malicious instructions are hidden in files, webpages, or tool responses that a model may treat as commands, which could lead to data being sent out.

OpenAI said the setting does not fully prevent prompt injection because harmful instructions can still appear in cached web content or uploaded files. However, it is meant to reduce the ways sensitive data could be sent out.

The move follows a broader pattern. Admins can turn off web search in Microsoft 365 Copilot, while Visual Studio Code enterprise policies can block access to agent tools and network domains.

That aligns with least-privilege guidance for generative AI, which treats prompt injection as an ongoing risk rather than a solved problem.

🔗 Source: OpenAI

Recent OpenAI developments