News

OpenAI has started delivering a new Lockdown Mode for eligible ChatGPT personal accounts intended to reduce the chances that sensitive information can be siphoned off through prompt injection techniques. The setting is aimed at individuals and organizations that manage confidential data and need more stringent protections. Lockdown Mode is being made available to signed-in users across Free, Go, Plus, and Pro tiers, as well as customers on self-serve ChatGPT Business plans.

According to OpenAI, Lockdown Mode is an optional, advanced security toggle that restricts a number of features and capabilities within OpenAI products that can reach the internet or external services. The company says the control is intended to shrink the avenues for data to be exfiltrated via prompt injection by curbing outbound network activity, noting that this comes at the cost of disabling or limiting some convenient features.

The mitigation focuses on hardening the platform against prompt-injection threats-an ongoing challenge across large language models (LLMs). Lockdown Mode extends existing sandboxing and other protections to specifically address URL-based exfiltration techniques by limiting outbound network requests that might forward sensitive content to attacker-controlled endpoints. The setting is not designed to eliminate prompt injections themselves, nor does it change how memory, file uploads, or conversation sharing operate; rather, it aims to remove possible routes that would allow stolen data to leave the system.

When enabled, Lockdown Mode turns off or restricts several capabilities, including:

• Live web browsing - limited to cached content only; real-time web requests are blocked.
• Image support - displaying images in standard responses and fetching images from the web are disabled.
• Deep research Agent mode - this agent-driven research capability is deactivated.
• Canvas networking - users cannot approve Canvas-generated code to make network calls.
• File downloads - downloading files for offline analysis is disallowed.

OpenAI cautions that Lockdown Mode is not intended for every user. It also clarified that Lockdown Mode and Developer Mode are mutually exclusive: enabling one will disable the other. While the company says Lockdown Mode significantly lowers the likelihood of prompt injection-based data exfiltration in ChatGPT and supported OpenAI products, it stops short of promising absolute protection. Residual risks may persist through enabled third-party apps, unexpected combinations of features, or newly discovered attack methods.

Moreover, Lockdown Mode does not eliminate all consequences of prompt-injection attacks. For example, a malicious directive embedded inside an uploaded file could still influence ChatGPT’s responses and produce incorrect or misleading outputs.

Alongside Lockdown Mode, OpenAI has introduced a session-management tool that lets users inspect active ChatGPT sessions and sign out of single or all sessions if they see signs of unauthorized access. The session list shows details such as the device and app used, approximate location, sign-in date and time, whether the device is marked as trusted, and whether it represents the current session.