Microsoft's AI web of the future had a major flaw.

Rather than simply extracting data from web pages, chatbots are evolving into real 'agents' capable of interpreting natural language queries to interact with online services. Through APIs or integrated tools, they can, for example, search for a flight, analyze a hotel offer, or compare prices.

This is the new trend for AI companies, everyone is getting into it, for example OpenAI with ChatGPT Agent, which can rely on integrations to plan a trip or perform online tasks.

Read ChatGPT makes a leap forward: with Agent, OpenAI's AI controls your PC and does everything for you.

It is now about making sure that websites and AI agents can understand each other better. During the spring Build conference, Microsoft therefore presented NLWeb with the idea of making it the « HTML of the agentic web », or more simply to allow AI to discuss directly with websites.

Unfortunately, barely launched into the internet, NLWeb presented a major security flaw. By examining the source code of the framework (whose code is open source), independent security researchers discovered that anyone could access sensitive files on a server simply by modifying a URL and without needing to log in. This major vulnerability represents a case of « path traversal » (« path traversal »), a well-known type of flaw but with this time far more severe consequences. A hacker could thus access files containing API keys (GPT-4, Claude, Gemini…) allowing them to use them fraudulently.

These keys are like the "brain" of the AI agent.

In the wrong hands, one can make the agent say anything (such as recommending fake medical treatments, for example), create a malicious clone, or inject harmful content. Moreover, files accessible through the vulnerability can also expose personal data, source code, or critical parameters.

In other words: it's a very serious flaw! Researchers warned Microsoft on May 28th, and a fix was applied two days later. So more fear than harm, but it's a warning for the publisher who wants to move very quickly (too quickly?) to establish itself in the agentive web.

The company did not want to assign a CVE — vulnerability identification number — even though it's standard practice. Perhaps because the usage remains limited for now. Microsoft explains that the affected code is not used in its products and assures that customers using the repository "are automatically protected ".

 

👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source : The Verge