OpenAI has unveiled Lockdown Mode, a security feature designed to protect against prompt injection attacks that attempt to extract sensitive data from AI systems. The feature adds strict constraints on how models process and relay information from external sources such as uploaded documents, URLs, and third-party tool outputs.
Prompt injection has been a persistent vulnerability in LLM-based applications, where attackers embed hidden instructions in data that trick the model into ignoring its system prompts or revealing confidential information. Lockdown Mode addresses this by isolating external data from the model instruction-following mechanisms and limiting what information can cross between contexts.
The feature is available for ChatGPT Enterprise and API customers, with granular controls that administrators can configure per application. OpenAI positioned Lockdown Mode as part of its broader push to make AI deployments enterprise-grade and compliant with data protection regulations.