OpenAI has announced the introduction of a new security feature called Lockdown Mode. This mode was designed in response to increasing risks associated with so-called prompt injection attacks, where attackers embed hidden instructions into websites or documents that can manipulate the chatbot's behavior and lead to the leakage of sensitive information.
Activating Lockdown Mode brings significant functional limitations in favor of higher security. Key measures include deactivating live web browsing, meaning the system will only access cached content. Furthermore, searching and displaying external images from the internet will be blocked, while the ability to generate new images remains. Deep research functionality and advanced agent mode will also be limited.
“Lockdown Mode is not for everyone. It is designed for individuals and organizations that handle sensitive data and require stricter protection against data exfiltration risks associated with prompt injection,” OpenAI stated in its announcement about the new feature.
Despite these measures, OpenAI warns that Lockdown Mode does not provide one hundred percent protection. The risk of chatbot manipulation still exists in cases where malicious instructions appear in cached web content or in files directly uploaded by the user. Attacks of this type can still affect the model's behavior or the accuracy of generated responses.
In the first wave, the new feature is being made available to users within self-service ChatGPT Business accounts and selected personal accounts that meet the specified criteria. With this step, the company is responding to criticism from security experts, who have long pointed out the vulnerability of large language models to external data sources.
The introduction of Lockdown Mode comes at a time when OpenAI is facing increasing pressure for transparency and data security, especially in the context of enterprise deployment of artificial intelligence technologies. The main goal of the tool is to reduce the likelihood of unintentional sharing of sensitive data during interaction with the chatbot, even if it comes at the cost of reduced user comfort and the timeliness of retrieved information.