- OpenAI rolled out Lockdown Mode to curb data leaks from prompt injections.
OpenAI announced that ChatGPT can now be run in Lockdown Mode, which disables the model's ability to make outbound API calls, web requests, or file system accesses during a session. The mode is optional and can be toggled per user or organization. OpenAI says the feature is designed to stop the model from inadvertently pulling in external content that might contain sensitive information. The company also updated its documentation to clarify the limits of the protection.
The change matters because many enterprises use ChatGPT for confidential tasks and worry about prompt‑injection tricks that could exfiltrate data. While Lockdown Mode reduces the attack surface, OpenAI admits it does not fully prevent malicious prompts from extracting stored context. Users will still need to vet inputs and monitor outputs.
In practice, the feature is another layer of defense, not a silver bullet, and its effectiveness will depend on how rigorously organizations enforce it.