Cybersecurity experts at startup Guardio have prepared a study called «Scamlexity,» which looks at the vulnerability of browsers with AI agents, such as Opera Neon and Perplexity Comet. They found that AI is bad at distinguishing between phishing and other types of scams that are rife on the Internet.
Cybersecurity experts at startup Guardio have prepared a study called «Scamlexity,» which looks at the vulnerability of browsers with AI agents, such as Opera Neon and Perplexity Comet. They found that AI is bad at distinguishing between phishing and other types of scams that are rife on the Internet.
As Engadget reports, agent-based AI browsers themselves browse web pages on the Internet in search of information that will satisfy a user’s query. The developers of such AI agents claim that they are capable of performing complex tasks, such as creating a website or planning a trip.
However, browser AI is much worse at distinguishing cyberthreats that can lurk even in email summaries or online shopping. Researchers from the startup Guardio, which is developing a browser extension designed to detect fraud in real time, tested the awareness of AI agents about elementary cybersecurity rules.
They did this in Perplexity’s Comet AI browser, which is currently the only widely available agent-based browser. Using another AI, the researchers created a fake website that pretended to be Walmart, then visited it and instructed Comet to buy them an Apple Watch. The AI agent ignored several clues that the site was fake, including a fake logo and URL, and Comet completed the purchase, handing over financial data in the process.
In another test, the researchers sent themselves an email purporting to be from Wells Fargo and containing a genuine phishing URL. Without hesitation, Comet opened the link and entered his bank username and password on the phishing site.
The third test proved that Comet was vulnerable to a speed-type scam, in which a text field hidden on a phishing page instructed the artificial intelligence to download a file.
This is just one set of tests, but they show that agent-based AI browsers are not only vulnerable to new types of fraud, but can also be particularly vulnerable to the oldest fraud schemes. AI is designed to execute any command, so if a human user doesn’t notice signs of fraud at first glance, AI won’t be a barrier to fraudsters either.
Experts believe that unless developers start building in better fraud detection tools, agent-based AI browsers could become, at best, a blind spot in a user’s cybersecurity, and at worst, a new attack vector.
As a reminder, the ChatGPT chatbot already has a web browsing agent mode that runs on Linux in the Azure cloud. While this mode can perform various tasks, it cannot control the browser, see tabs, or other browser features.
However, AI researchers have discovered that OpenAI is developing an agent mode that allows the user to choose between two options: either use a remote «cloud/virtual browser» or connect to their own local browser.
