OpenAI Bolsters ChatGPT Security Against Prompt Injection
OpenAI has introduced two new advanced security features for ChatGPT to combat the rising threat of prompt injection attacks, which occur when AI systems interact with websites and external applications. These attacks can allow malicious actors to manipulate the AI into performing harmful actions or exposing confidential data. Both new features enhance the platform's existing defenses, including its sandbox environment, URL data leak prevention, and other multi-layered security protocols.

The first feature is Lockdown Mode, an optional, advanced setting designed for users with stringent security requirements, rather than the average user. It significantly lowers the risk of data leaks by imposing strict controls on how ChatGPT communicates with external systems. This mode proactively disables or restricts tools that are potential attack vectors. For instance, the web browsing feature will only access cached content, and functionalities without robust security assurances will be turned off entirely.
Currently, this mode is available for ChatGPT Enterprise, Education, Healthcare, and Teacher editions. Administrators can enable it by creating specific roles in the dashboard settings, where they can also define precisely which external applications and actions are permitted while Lockdown Mode is active. OpenAI plans to make this mode available to consumer users in the coming months. Furthermore, the Compliance API Logs Platform assists administrators in conducting regulatory audits of application usage and data sharing practices.
The second measure involves consistently tagging higher-risk functions within ChatGPT, ChatGPT Atlas, and Codex with an **"Elevated Risk"** label, standardizing how potential threats are communicated to users.
OpenAI explained that certain internet-connected capabilities, while increasing the AI's usefulness, still present risks not fully mitigated by current industry safeguards. This label provides clear visibility into these risks, empowering users to make informed decisions about whether to use the features, particularly in situations involving sensitive information. For example, when developers enable internet access for Codex, the label will appear on the interface, detailing the changes it brings, potential dangers, appropriate use cases, and recommendations for reducing risk.
Related article
Cyberspace Administration of China mandates tagging of AI-generated and fictional short videos
The Cyberspace Administration of China has rolled out a comprehensive plan to standardize short video content labeling, mandating that platforms offer six required tags—including "AI-generated content"—ushering in a new era of mandatory transparency
DeepL, renowned for text translation, now targets voice translation
DeepL, a translation company best known for its text-based tools, has launched a voice-to-voice translation suite today that addresses scenarios such as meetings, mobile and web conversations, and group discussions for frontline workers through custo
Talat’s AI meeting notes live on your device, not the cloud
Granola, the AI-powered notetaking app valued at $250 million, has gained traction among tech founders and venture capitalists. But one developer sees demand for a more private, fully local alternative available for a one-time fee with no subscriptio
Related Special Topic Recommendations
Comments (0)
0/500
OpenAI has introduced two new advanced security features for ChatGPT to combat the rising threat of prompt injection attacks, which occur when AI systems interact with websites and external applications. These attacks can allow malicious actors to manipulate the AI into performing harmful actions or exposing confidential data. Both new features enhance the platform's existing defenses, including its sandbox environment, URL data leak prevention, and other multi-layered security protocols.

The first feature is Lockdown Mode, an optional, advanced setting designed for users with stringent security requirements, rather than the average user. It significantly lowers the risk of data leaks by imposing strict controls on how ChatGPT communicates with external systems. This mode proactively disables or restricts tools that are potential attack vectors. For instance, the web browsing feature will only access cached content, and functionalities without robust security assurances will be turned off entirely.
Currently, this mode is available for ChatGPT Enterprise, Education, Healthcare, and Teacher editions. Administrators can enable it by creating specific roles in the dashboard settings, where they can also define precisely which external applications and actions are permitted while Lockdown Mode is active. OpenAI plans to make this mode available to consumer users in the coming months. Furthermore, the Compliance API Logs Platform assists administrators in conducting regulatory audits of application usage and data sharing practices.
The second measure involves consistently tagging higher-risk functions within ChatGPT, ChatGPT Atlas, and Codex with an **"Elevated Risk"** label, standardizing how potential threats are communicated to users.
OpenAI explained that certain internet-connected capabilities, while increasing the AI's usefulness, still present risks not fully mitigated by current industry safeguards. This label provides clear visibility into these risks, empowering users to make informed decisions about whether to use the features, particularly in situations involving sensitive information. For example, when developers enable internet access for Codex, the label will appear on the interface, detailing the changes it brings, potential dangers, appropriate use cases, and recommendations for reducing risk.
Cyberspace Administration of China mandates tagging of AI-generated and fictional short videos
The Cyberspace Administration of China has rolled out a comprehensive plan to standardize short video content labeling, mandating that platforms offer six required tags—including "AI-generated content"—ushering in a new era of mandatory transparency
DeepL, renowned for text translation, now targets voice translation
DeepL, a translation company best known for its text-based tools, has launched a voice-to-voice translation suite today that addresses scenarios such as meetings, mobile and web conversations, and group discussions for frontline workers through custo
Talat’s AI meeting notes live on your device, not the cloud
Granola, the AI-powered notetaking app valued at $250 million, has gained traction among tech founders and venture capitalists. But one developer sees demand for a more private, fully local alternative available for a one-time fee with no subscriptio





Home






