OpenAI Secures ChatGPT with New Lock Mode

OpenAI has recently activated a new security feature for its ChatGPT users, called Lock Mode. This feature, initially mentioned in February, is now available for all eligible personal accounts, including Free, Go, Plus, Pro versions, as well as for self-service ChatGPT Business accounts.

Lock Mode has been designed to counter the latest stage of prompt injection attacks by limiting outgoing network requests that could potentially transfer sensitive data to an attacker. While this measure cannot prevent the occurrence of prompt injections in the content processed by ChatGPT, it aims to prevent the exploitation of these injections to steal data.

One of the main challenges faced by large language model (LLM) systems is the "Deadly Trio." This trio manifests when the system has access to private data, is exposed to unreliable content, and has a means to transmit this data to an attacker. To address this issue, it is crucial to cut off one of these three elements. Lock Mode specifically targets the data exfiltration element, using deterministic mechanisms that do not rely on AI systems, which can be vulnerable to sophisticated attacks.

However, the introduction of Lock Mode also highlights that, in its default settings, ChatGPT does not provide complete protection against determined data exfiltration attacks. Therefore, this new feature represents an important step in enhancing user security against increasingly sophisticated cyberattacks.