OpenAI deploys Lockdown Mode to protect ChatGPT from prompt injection attacks

OpenAI has launched the new Blocking Mode, which is now available in Spain, which limits access to the website and external services to reduce the risk of attacks of ‘prompts’ injection.

Lockdown Mode is characterized as a new security option that limits the tools and capabilities of OpenAI products that can connect to external services or the web.

The objective is reduce data exfiltration due to prompt injection attacks (when text instructions are manipulated to try to trick the AI ​​into ignoring its original rules), by limiting outgoing network requests by blocking or some of ChatGPT’s useful features.

OpenAI, from its website, indicates that this new blocking mode is not aimed at all users, but rather to organizations or people who often work with sensitive data and need the necessary protection to limit the risk of this data being exfiltrated, that is, ending up transferred to an environment controlled by the attacker.

What this mode does is apply a series of protections at the model, product and system level. This includes everything from process isolation to protections against URL-based data exfiltration, to policy monitoring and enforcement as well as enterprise controls such as role-based access and audit logs.

The creators of ChatGPT warn that the new protection mode does not prevent prompt injections from appearing in content processed by ChatGPT. Among these cases is the possibility that they appear in the cached content of a website or in a file that the user has uploaded.

That is, it is not a shield that eliminates all attacks, but what it does do is disable those actions or functions that could be used to trick the AI ​​with a prompt injection attack.

OpenAI has indicated that when Lockdown Mode is active real-time web browsing, support for images is disableddeep investigation function, agent mode, Canvas and file download. It clarifies that it does not interfere with memory, file uploads, the ability to share a conversation or that they are used to improve the models.

This new ChatGPT update reaches the personal accounts of all usersincluding the free version, Go, Plus and Pro, as well as ChatGPT Business accounts.

It is part of a global deployment that is being carried out these days and to which can be accessed from the ‘app’ settings to find the Lock Mode in the security section.