The one thing I learned is that you can actually get Malware without having to click the ad. It got to the point where even the FBI themselves recommend users to use an ad-blocker because of this. Although it's not widespread and only targets high profile users, it's scary to think about that it's even possible to begin with and crazy how it's not patched.

Here is a video that explains it.

Otherwise, good stuff :)

One thing that stood out while building OmniGuard was how easily deceptive elements can blend into an otherwise normal-looking webpage.

A countdown timer, a misleading button, or a tiny pre-selected checkbox might not look like much on its own, but these little design decisions can completely change what a user ends up doing.

We built OmniGuard to make these tricks easier to spot, but I'm sure there are plenty of dark patterns we haven't thought of yet.

What's one website or app experience that made you feel genuinely tricked? I'd love to hear some real examples! 👀

Great piece of technology you people built there!

What's the rate of false positives? For example, high-demand concert tickets: Let's say a very famous pop star comes to town and the local ticket vendor only allows tickets to be in the cart for ~10 minutes, so black market resellers don't hoard them to create artificial scarcity and inflate the price later on. Will such a countdown be flagged? Not saying that it's a good thing per se, it does create urgency, but it's actually a "necessary evil" to protect legitimate fans from illegitimate resellers and it actually does function as information to the user.

Same goes for false negatives: A page that looks like it would have a working download button, because the page content implies a download and the button looks and feels legit enough to be the download button, although it's an ad - will this be flagged?

Although you're not using a database, won't Google store these screenshots anyways when you send them to Gemini, or am I missing something? Or perhaps there's a way to redact certain parts of the screen beforehand?

And now for the opinion part: The best way to get rid of these dark patterns is by having functioning consumer protection laws and actually fining the living heck out of companies that don't oblige. I know that such laws are next to impossible to install everywhere, let alone enforce, but I like to dream...

While reading through this, my first concern was about privacy and whether the extension captures the entire screen or potentially sensitive information. I was relieved to see that you've addressed this by limiting screenshots to the active viewport and using a stateless, zero-database architecture.

That said, one thing I'd be curious about is how screenshot data is handled during transmission and by Gemini itself. Even if your backend doesn't store anything, the screenshots could potentially contain sensitive information visible in the viewport. Clarity around API-side data retention, logging, and processing would make the privacy story even stronger.

Also, I find the distinction between traditional ad blockers and AI-powered visual detection particularly interesting. Tools like Brave already do a great job with trackers and malicious content, but detecting psychological manipulation at the UI level is a different challenge altogether.

Really interesting project! Great work by the team. 👏

Brilliant initiative! The web should earn our trust, not manipulate us. Could tools like OmniGuard become the new standard for ethical design?

This is a great example of how AI can make the web more user-friendly. Dark patterns don't just hurt user trust; they also damage the overall user experience. I really like the idea of using visual AI to identify deceptive design elements in real time. Technology should help users make informed decisions, not manipulate them. Great work!

The fake timer example makes me want a separate uncertain result. From one screenshot I can see urgency language, but I can’t tell whether the deadline is genuine or whether the timer resets on another visit. Would you consider showing the visual evidence and confidence beside the warning, instead of treating both cases as equally established?

Another tip for anyone dealing with ads, use AdGuard (the green shield), honestly the best blocker imo.

The worst one I've seen recently: a 'free' PDF converter that let you upload, process, and preview the file — and then demanded a subscription to download it. The preview was the trap. That's a variant of forced continuity that I'd love to see OmniGuard catch.

Dark patterns live in desktop installers too — the fake 'download now' buttons and pre-checked bundle offers show up in Electron apps and their update prompts, where there's no DOM to inspect, only pixels, so a visual approach like this is exactly the right shape. Pairing the detector with an on-device model would close the screenshot-privacy gap even further at the cost of some latency. And the Shadow DOM injection is a nice detail — it keeps host-page CSS from bleeding into your overlay.