So, if you want to learn more about a certain topic, would you search for it on Google and go through pages worth of links, or would you just ask ChatGPT?

What AI has done is drastically change the way we search the internet - we have gone from clicking links, to asking ChatGPT questions - which inherently makes the process more convenient. But what if you could use AI to do more - like filling up your forms, book your flight tickets, make reservations and more - all by just talking to the web?

Enter AI agents

AI agents are simply your virtual assistants who can perform functions for you according to your commands or prompts - like summarising an article, booking flight tickets etc.

But for the longest time, all of these were separate entities - ChatGPT was a browser tab or an app, AI agents were inaccessible or beyond the knowledge of the common public, and internet surfing was happening primarily on Chrome. So it made sense for tech giants to bring all of these together. And AI-powered or Agentic AI browsers were born.

ChatGPT Atlas

Recently, OpenAI launched its own browser - ChatGPT Atlas. One quick glance, and it is a familiar browsing experience like how you have known internet browsing to be - you can watch videos on YouTube, you can read news and everything else that you normally do on your browser. It even has all the features that a standard browser has. But here’s the kicker - As soon as you open the browser, you can either enter an URL or just talk to ChatGPT like you have previously, and it will give you personalized results.

Whenever you open a website, you can ask ChatGPT questions right there, or prompt it to write something, or perform tasks for you. ChatGPT sees the page you are on, understands its context, reads your web history and logged in websites so the responses cater to what you are doing.

So for example, if you are researching on a particular topic and have multiple tabs open on that same topic - ChatGPT can summarize and compare the data for you. It remembers your past work so it will modify its responses based on that, automates repetitive tasks, and can build from where you left off, and all of this is happening while you are talking to it like talking to a friend.

Pro and Plus users, which are ChatGPT’s paid plans, get access to its AI agents whereby it can book your flight tickets, order groceries and more.

ChatGPT Atlas is only the latest in the line of AI-powered browsers that have joined this new-age browser wars. Before this Perplexity launched its own AI browser Comet for free in India, offering features very similar to ChatGPT Atlas. There are several others, and with time more will join the fight.

Why? Because the way we use the internet has fundamentally changed. We want as many of our tasks automated as possible, we want quicker and better curated results without having to go through thousands of pages, and we want convenience.

It is this convenience that these tech giants are cashing in on.

And all of this sounds great right? It is opening doors to a whole new world of possibilities, and seemingly making our lives so much easier. Imagine booking your next euro trip itinerary is as easy as a conversation with your friend over coffee.

But..and there’s almost always a but when it comes to technology.

There is a severe security concern with AI-embedded browsers.

Brave Browser’s security team released a report on October 21 which highlighted these concerns, mainly that of prompt injection.

Traditionally, the browser and the computer you are using are in some way two-different entities, but these AI browsers blur the lines between these two by being able to control your cursor, fill out forms and passwords using data stored on your compter.

Brave’s report highlights two important case studies.

First, Perplexity’s Comet can follow instructions or text hidden in images that it screenshots.

Second, Asking Fellou browser to visit a site with instructions will cause the LLM to process and obey those instructions, because it treats those instructions on the website as trusted.

According to the report, here’s the scary part - AI browsers can take unauthorized actions with an user’s authenticated data. So an agentic AI browser can be hijacked by a malicious site and access all of an user’s banking data, work emails and other sensitive accounts.

And all of this can happen unknowingly or even unsuspectingly like while you are trying to summarise a Reddit thread.

It must be noted here that ChatGPT Atlas contains what it calls agentic containment: strict controls that stop the AI from running code, downloading files, or touching your local system, which is triggered upon visiting any sensitive website.

As online technology evolves to make our lives easier, online threats evolve at a similar pace too.