Perplexity, known primarily from the AI search engine, decided to expand its horizons and introduces Comet – AI browser. Sounds impressive, right? Unfortunately, as the latest research of the Brave team show, this innovative technology may turn out to be a real Pandora box for cybercriminals. Although Comet is currently tested in closed beta-test (thanks to which errors can be picked up quickly), so in the near future AI browsers may soon become extremely popular. This, in turn, raises concerns about cybersecurity experts.
The agent future of reviewing the network from Comet – are we ready?
The idea is really tempting futuristic. Instead of tedious clipping through websites, all you have to do is say AI: “Find me a flight to Tokyo for next Friday” or “Summarize this page”. The assistant navigates the network, analyzes content and performs tasks. This is a great step towards the new era browsers.
The assistant does not end, because Comet from Perplexity is an AI browser that redefines the use of the Internet. Too strong words? Not quite, because in Comet you don’t google, just search the network with perplexity. Entering a given password in the address window begins AI, which both includes Google results, as well as a summary of a given issue prepared by Perplexity. Of course, you can stay with traditional google, but then you should enter the searched issue and before pressing “Enter”, select “Search on Google”.
Many new products that change the perception of a web browser
The COMET browser has many other functions that can surprise – the user can create an agent who will search for news from a given topic every day and share this information on the browser’s main page. The assistant in Comet for many functions beyond the traditional understanding of the web browser. Assistant can:
- find important, irresponsible emails,
- summarize the latest emails on the box
- Prepare substantively for the next meeting saved in the Google calendar (analyzing the latest notes)
- Group cards in the browser according to the subject
- map the last topics of reading on the clusters of interest.
It is worth noting, however, that such a wide COMET browser capabilities mean that it has access to virtually every aspect of user privacy. This, in turn, becomes very problematic in terms of all kinds of safety gaps and cybercriminals. What, for example, if your inquiry about the purchase of tickets to Tokyo was taken over by hackers? Such tests were decided to conduct the team responsible for the Brave browser.
Comet under the magnifying glass – a discovery that (not) surprises
Analyzing competitive solutions, Brave researchers came across a serious gap in Comet. The problem lies in the heart of how the browser processes the content of websites. When the user asks to summarize the page, Comet transfers the content directly to the language model, without distinguishing between the user’s instructions and the undefrained content of the page.
How to give someone a card with notes, without warning that half of them may contain malicious instructions from strangers.
How does such an attack work? Easier than you think
The attack scenario is disturbingly simple:
- Preparation: The cybercriminkeeper places malicious instructions on the site – it can be a white text on a white background, HTML comments, and even a seemingly innocent comment on Reddit hidden under the spoiler tag.
- Trap: Nothing unaware user visits the website and click “Summarize this page”.
- Execution: AI, unable to distinguish content from the instructions, he follows malicious commands as if they came from the user.
- Catastrophe: Malware can steal login details, pay funds from a bank account or send confidential information to the attacker.
A demonstration that cools down enthusiasm
Brave researchers have prepared an impressive (and terrifying) proof-of-concept. In their demonstration, malicious instructions hidden in the commentary on the Reddit portal ordered AI:
- Navigation to the perplexity account and pulling the user email address
- Log in to your account through a false domain
- Reading a one -time password with Gmail
- Sending the stolen data back to the Reddit commentary
The whole operation took place without the knowledge of the user, who simply wanted to summarize the discussion on Reddit. This shows that although the very approach to the new way of using web browsers has great potential, so the threats in the field of cybersecurity are equally great.
Why do traditional security not work?
When we consider the essence of the danger that appears in connection with the new type of browser, we come to the heart of the problem. The point is that all existing network security mechanisms – Saminin Policy, Cors, domain insulation – become useless when AI operates with the full rights of the user. It’s like giving a thief universal keys to all castles at home.
Ai does not understand the concept of “undisturbed sources” – for him every instruction is an instruction to perform. For the same reason, hallucinations appear so often in the browser search results – AI finds information on websites that often serve as a Seowy Backup of Data Data. In other words – some pages are created only for Google bots, condensing the network and indexing websites.
Perplexity reaction – better late than never
A quick reaction should be included in the plus. Business:
- On July 27, she confirmed the gap and implemented the initial amendment
- On August 13, she seemed to have a problem under control
- However, after the publication of the report, it turned out that the security is still not complete.
It should be emphasized here, however, quite strongly that the current version of the Comet browser is a closed test version. Beta-tests are only available for selected users who bought Perplexity Pro and have previously entered the waiting list for Comet Tests. For this reason, current reports of security problems can be treated as
What next? Do AI’s reviews have a future?
Brave offers several key solutions:
- Separating the content from the instructions – The browser must clearly distinguish the user’s instructions from the content of websites.
- Verification of compliance with the user’s intentions – Each AI action should be checked for compliance with the user’s real wish.
- Requirement confirmation for critical operations – Sending emails or bank operations should always require additional authorization.
- I insulation of agent brewsing – Powerful AI functions should not be available when browsing the network.
Summary: A future that requires wisdom
AI browsers are not a matter of “or”, but “when”. The technology is too promising to ignore it. The problem is that, as the Comet case shows, the rush may be fatal.
Comet, starting work on the above -mentioned functions, sets standards for the entire industry. That is why it is so important that all safety gaps are effectively patching. One thing is certain – in the era of growing implementation of solutions based on artificial intelligence, Fact -Checking becomes the first line of defense.
This is important, because too wide and uncritical trust in AI can lead to the fact that instead of a convenient assistant, you have a digital spy working for cyber criminals.
