I Let Google’s ‘Auto Browse’ AI Agent Take Over Chrome. It Didn’t Fairly Click on

After I was lastly capable of experiment with Auto Browse (for actual this time), I took Google’s options of digital chores as my start line and picked on-line duties that might be useful in my very own life.

At any time when interacting with generative AI instruments, a wholesome sense of skepticism—and warning—is crucial. Google even features a disclaimer baked into its Gemini chatbot reminding customers that it makes errors. The Auto Browse instrument goes a step additional. “Use Gemini fastidiously and take management if wanted,” reads persistent textual content that exhibits within the chatbot sidebar each time Auto Browse is working. “You might be answerable for Gemini’s actions throughout duties.”

Earlier than you attempt it out, you additionally want to consider the safety dangers related to this sort of automation. Generative AI instruments are susceptible to being compromised via immediate injection assaults on malicious web sites. These assaults try to divert the bot from its activity. The potential vulnerabilities in Google’s Auto Browse haven’t been absolutely examined by exterior researchers, however the dangers could also be much like different AI instruments that take management of your laptop.

As well as, take further warning in the event you’re utilizing Auto Browse to make purchases. Google has safeguards in place that flag sure actions, like shopping for stuff or posting on social media, as delicate and in want of person approval to proceed. Nonetheless, I used to be uncertain how the bot would behave and anxious concerning the havoc it might doubtlessly wreak with my bank card, to say nothing of handing over monetary data to it within the first place.

Right here’s the primary immediate I despatched it, card in hand:

I need to e book two tickets to the SF symphony tonight. I do not need to pay for orchestra seating, however the tickets do not should be the most cost effective ones out there. Please decide the 2 seats subsequent to an aisle.

It’s a bit weird to observe Google’s AI agent click on round within the tab. First, I noticed it use Gemini 3, Google’s newest mannequin, to strategize and outline targets, like getting two aisle seats on the symphony, within the sidebar textual content field for a couple of seconds. This course of seems to be much like a chatbot utilizing a “reasoning” mannequin, speaking via the steps it would take earlier than transferring ahead. Then, the press begins. Every step the bot takes as a part of a activity is logged for customers.

Auto Browse’s skill to carry out multistep duties with out getting sidetracked was noticeably higher than comparable agent instruments that I examined final 12 months. It navigated to the right web site, selected the best efficiency, and clicked on a number of seat sections to gauge availability. The whole lot listed within the log seemed to be what it truly executed.

After a few minutes of engaged on the symphony tickets, the bot stopped clicking. I acquired a notification to take over and press the Order Now button. At a look, the AI instrument had seemingly delivered what I’d requested for, and moderately shortly.

But when I had unquestioningly ordered the 2 seats Auto Browse selected for a date on the symphony, the evening would more than likely have ended with my boyfriend making me sleep on the sofa.