OpenAI announced the introduction of Lockdown Mode and standardized Elevated Risk labels in ChatGPT to help organizations and individual users mitigate emerging security threats, particularly prompt injection attacks and risks associated with AI systems interacting with connected apps and the web. These enhancements build on OpenAI’s multi-layered safety protections and offer clearer guidance and stronger controls for high-risk use cases.
OpenAI explained that as AI systems increasingly handle complex tasks involving external connections and sensitive workflows, potential adversarial risks such as prompt injection attacks – where malicious actors attempt to mislead AI into executing harmful instructions or extracting sensitive data – require advanced mitigation strategies. Lockdown Mode and Elevated Risk labels are designed to provide users with stronger safeguards and greater situational awareness when working with potentially vulnerable capabilities.
Lockdown Mode is an optional advanced security setting intended for highly security-conscious users and organizations, such as executives, cybersecurity teams, and other profiles at elevated risk of targeted attacks. This mode tightly constrains how ChatGPT can interact with external systems. Under Lockdown Mode, certain tools and capabilities that might expose data risk — including unrestricted web access – are deterministically disabled or limited to minimize opportunities for prompt injection-based data exfiltration. For example, live web browsing is restricted to cached content so that no live network requests leave OpenAI’s controlled environment, reducing the potential for sensitive information to be transmitted to unauthorized parties.
Also Read: QuiX Quantum and Artilux Establish Strategic Collaboration to Advance Energy-Efficient Photonic Quantum Computing
Lockdown Mode is available immediately for customers on ChatGPT Enterprise, ChatGPT Edu, ChatGPT for Healthcare, and ChatGPT for Teachers plans. Workspace administrators can enable this feature through their Workspace Settings by assigning a dedicated Lockdown Mode role. Admins also retain granular control over which trusted apps and specific actions within those apps are permitted for users operating under this mode. OpenAI plans to make Lockdown Mode available to consumer and team plan users in the coming months.
In addition to Lockdown Mode, OpenAI introduced Elevated Risk labels across ChatGPT, ChatGPT Atlas, and Codex to help users make informed choices about features that may introduce additional security risk when connecting AI products to networks, apps, or externally invoked capabilities. These labels provide consistent in-product guidance by highlighting when a capability carries higher potential for data exposure and explaining what changes, risks, and trade-offs are involved. For instance, in Codex, enabling network access for web-based documentation or actions is accompanied by an Elevated Risk label that clarifies both its utility and the associated security considerations.
OpenAI emphasized its continuing investment in safety and security safeguards, noting that over time Elevated Risk labels may evolve or be retired once certain risks are sufficiently mitigated through ongoing advancements. The company also plans to maintain oversight of which features are labeled in order to best communicate relevant risk information to users as AI capabilities and security protections progress.
SOURCE: OpenAI