To reduce the risk of sensitive data being maliciously stolen, OpenAI has recently introduced a new optional security setting called "Incident Mode" for all logged-in users. This mode covers different account types and workspaces, and users and administrators can choose to enable it on their own. Once this feature is activated, ChatGPT will actively restrict its powerful real-time internet access, in-depth research, and agent modes. This measure aims to build a security firewall against prompt injection attacks by limiting the model's ability to initiate external network requests.

ChatGPT OpenAI Artificial Intelligence (1)

Blocking Internet Functionality to Plug Data Leaks

In "Incident Mode," ChatGPT's real-time web browsing capability will be strictly limited, and the system will only be allowed to access cached web content, which may result in incomplete or outdated search results. At the same time, the model will also be unable to download any files from the internet for data analysis, and the code generated by Canvas will also be prohibited from initiating internet access. However, local-related functions such as images uploaded by users, images generated using tools, and conversation memory will not be affected by this mode.

Returning Security Choices to Users

What is known as a prompt injection attack involves attackers hiding malicious instructions within web pages or documents, luring AI to ignore existing rules and thereby sending users' sensitive information to external servers. OpenAI admits that "Incident Mode" cannot completely eliminate such attacks, as malicious code could still be hidden in cached web pages or user-uploaded files. Faced with the reality that stronger AI capabilities mean a larger attack surface, OpenAI chooses to return the trade-off between "stronger functionality" and "lower data leakage risk" to the users themselves.