Ever since ChatGPT was released, many users have wanted to try this type of productive AI system. One of the easiest ways to do this was with browser extensions, and while that gives you all sorts of options, these extensions turned out to be one thing above all else: a threat to the security of our computers.

Above all and very dangerous. This is pointed out by the cybersecurity experts at Kolide who have conducted a study where they have made it clear that there are many extensions designed to steal data directly from us in this new wave of IA extensions, this is by no means new. . Others have evolved rapidly and have the copy-paste privacy terms of any other service, and others are artificial intelligence experiments from these true, well-known and respected companies.

Do not use sensitive or private data. The problem is that users all over the world install these extensions to “edit” their features, and in doing so they often bring sensitive data where it shouldn’t. Risk is a thing, especially for professional users who take advantage of these extensions at work.

Data stealing extensions. In March, digital identity protection company Guardio reported one of the first cases discovered. A Chrome extension called “Quick Access to Chat GPT” was stealing users’ Facebook accounts, as well as a list of all cookies stored in the browser, including those related to security tokens.

Google can’t cope. Although it only took the company a week to detect and remove it, the damage was already done: 2,000 users a day had downloaded and used it. Something similar has happened with many more extensions that promise ChatGPT from the browser, for example. By the way, something similar happens with mobile phones, where users pay to use this type of tool.

Various security threats. Among the threats are not only the theft of sensitive data, but also security problems of even official extensions – OpenAI detected a bug on its platform. There’s also a problem with copyright: Produced texts, codes or images can become partial reproductions of copyrighted works, leading to significant legal issues. Stack Overflow moderators went on strike over how the site is overrun with AI-generated code.

Quick injection for evil. The operation of ChatGPT triggered another type of cyberattack called ‘rapid injection’. The idea is simple: try to make the AI ​​model do things it shouldn’t do with a certain prompt. This is how Microsoft Edge admitted its “robot laws” to us, but attacks that can be carried out using this method can go much further and be more dangerous. A developer named Simon Willison stated in May. How a combination of ChatGPT Plus plugins can be used to extract private data.

How to avoid. The current functioning of LLMs makes them vulnerable to such attacks, and there seems to be no solution on the horizon, according to Kolide experts. In fact, they explain that the only way to expose ourselves is to not make things easy for these extensions (or ChatGPT itself): you don’t need to enter custom data on these platforms and link them to private data sources either. For example, it is not a good idea to give ChatGPT data about our financial situation in order to give us some form of financial advice, or ask it to provide any preliminary medical diagnosis, symptoms, or private and factual clinic. data as reference.

And what should companies do? Some already know very clearly that ChatGPT does not know how to keep secrets, and that includes those with confidential information, so they directly ban or significantly limit their employees’ access to ChatGPT. The advice to all is to at least educate and train their employees on the use of these extensions (and the risks involved in doing so), and to create “whitelists” of verified and available extensions.

Image | Denny Muller

on Xataka | The dangerous risk of missing the Apple and AI train