AI Found a $250K Chrome Hole in Six Days and I've Had a Ticket Open for Six Months
Oh good. Good. An AI just chained together three separate holes in the fence inside Chrome, built a full exploit that pops a Windows box just from loading a webpage, and did it in six days. Six. Days.
I've been waiting six days for the Shepherds to approve my request to block a single suspicious IP address. But sure, let's let the robots have all the fun.
Here's what actually happened. The researchers at XBOW pointed OpenAI's GPT-5.6 Sol model at Chrome like a sheepdog with a purpose and a caffeine drip. The AI agents found three separate zero-days, holes in the fence nobody knew existed, and stitched them together into one clean chain. Visit a webpage. Get owned. That's it. That's the whole attack. The Flock doesn't even have to click anything suspicious. They just have to exist near a browser tab.
Google confirmed all three bugs, paid out the full $250,000 bounty, and quietly patched everything. Good for them. Genuinely. But let's not lose the plot here.
The part that's keeping me awake, and I was already awake, is that XBOW used a publicly available model. This wasn't some classified government AI locked in a bunker. This was off-the-shelf. The Coyotes are reading the same product page you are. The only difference is they have a goal and the Lambs have a YouTube autoplay queue.
We are now in a world where the cost of sophisticated exploit research just dropped to whatever an API subscription costs. The six-day timeline is the real headline. Not six months. Not a red team engagement billed at $400 an hour. Six days and a credit card.
I need a nap and a new career.
Remediation
Look, I know. You're tired too. Here's the short list:
Keep Chrome updated. These specific holes are patched. Shearing already happened. Don't be the one still running the version from March.
Enable auto-updates everywhere. Yes, everywhere. Yes, including that one machine in the corner nobody talks about.
Isolate your browsing environments where it matters. Sensitive systems should not be casually surfing the Sky Pasture on their lunch break.
Talk to your Shepherds about AI-assisted threat modeling. They won't understand it, but say it loudly in a meeting and attach a dollar figure. $250,000 tends to wake them up briefly.
Assume the Flock will load the page. They always load the page.
Anyway, my ticket queue just hit 200. I'm going to go lie down in a field somewhere and think about goats.
Original Report: https://thehackernews.com/2026/10/researchers-used-public-openai-model-to.html