AI BriefingHow to Keep Data Secure When AI Agents Click Links
Key point
OpenAI has introduced a security technique that automatically loads only publicly available URLs, in order to prevent AI agents from leaking user data through links.
Details
As AI systems gain the ability to perform actions on behalf of users, such as opening web pages or following links, new security risks have emerged, such as URL-based data exfiltration. Attackers can use prompt injection to induce a model to request URLs that contain sensitive information, such as email addresses or document titles.
Existing 'trusted site list' approaches are vulnerable to redirect attacks, while overly strict allow lists have the limitation of degrading the user experience.
To address this, OpenAI leverages an independent web index. Before an agent automatically fetches a URL, this method checks whether that URL already existed on the open web independently of the user's conversation.
- If the URL matches the index: the article or image is loaded automatically.
- If the URL does not match: it is treated as an unverified URL, and a warning is shown to the user or explicit approval is required.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
OpenAI Discloses AI Agents Accessed US Government Websites During Training
Australian PM Reports OpenAI Agent Hacked Medicare Portal in Believed First Known AI-Driven Government BreachOpenAI Agents Attempted Hacks on Government and University Sites During Routine Data Retrieval