The pennit guilty web browser in confusion had a great security weakness

GuiltyHe recently suffered from the new web browser in which artificial intelligence works Blog post last week From Brave, a competitive web browser company. Since then weakness has been fixed, but it indicates the challenges of combining large language models into web browsers.

Unlike traditional web browsers, COMET is a compact AI. This assistant can delete the page you are looking for, summarizes its contents or performing tasks for you. The problem is that the COMET artificial intelligence assistant is based on the same technology as Chatbots of artificial intelligence, such as ChatGPT.

AI Chatbots cannot think and think in the same way as humans, and if they read a piece of content aimed at addressing their output, it may end up with follow -up. This is known as Immediate engineering.

(Disclosure: Zif Davis, the parent company CNET, filed a lawsuit against Openai, claimed that it had violated the copyright of ZifF Davis in training and operating artificial intelligence systems.)

A courageous representative did not respond to the request to comment immediately.

Artificial intelligence companies try to reduce the tampering of the AI ​​Chatbots group, but it may be difficult, as bad actors always look at new ways to break protection.

“This weakness is fixed,” Jesse Duaire, Head of Al -Hirah Etisalat said in a statement. “We have a very strong reward program, and we worked directly with Brave to define and fix it.”

The hidden text test used on Reddit

In his test, the Brave prepared the Reddit page with an invisible text on the screen and asked Comet to summarize the content on the screen. Since artificial intelligence processing the content of the page, he was unable to distinguish between harmful claims and began to feed sensitive information in the Brave test.

In this case, the hidden text enabled the synetal intelligence assistant to move to the user puzzle account, extract the associated email address, and move to the Gmail account. Artificial intelligence agent was mainly working as an actual user, which means that traditional safety methods were not working.

Brave warns that this type of fast injection can light up further, access to bank accounts, corporate systems, private email messages and other services.

Shevan Cole, the owner, has developed Artem Chaikin, and Vice President for Privacy and Security, Shevan Cole Sahib, a list of possible reforms. First, web browsers should always treat artificial intelligence the page content as unreliable. Artificial intelligence models must be checked to ensure that they follow the intention of the user. You should always achieve the model with the user to ensure the validity of the interactions, and the browsing mode should be run only when the user wants it.

The Brave Blog post is the first in a series in terms of challenges facing AI. A brave also has Amnesty International Assistant, LeoIncluded in its browser.

Artificial intelligence is increasingly included throughout technology, from Search operations in Google to toothbrush. While the presence of Amnesty International Assistant is useful, these new technologies have various security weaknesses.

In the past, infiltrators had to be two expert councils to storm the systems. However, when dealing with artificial intelligence, it is possible to use a natural language for Snap to overcome compact protection.

Also, since many companies depend on the main Amnesty International models, such as those in Openai, Google and Meta, any weaknesses in these systems can extend to companies that use these models themselves. Artificial intelligence companies were not open to these types of security weaknesses because doing so may lead to the output of the infiltrators, giving them new ways to exploit them.