Turns Out You Don’t Need the Most Powerful AI Models to Cause a Major Cybersecurity Mess

Fewer than 20 prompts. Under 24 hours. That’s what it took a cybersecurity firm called A Security to find a zero-click remote code execution bug in Zoom, using AI models anyone can access right now.

Not a frontier model behind a waitlist. Not some restricted enterprise tier with a six-figure contract attached. Publicly available tools, pointed in the right direction by researchers who knew what they were looking for.

The bug itself is about as bad as bugs get

Zero-click means the target does nothing wrong. No malicious link, no sketchy download, no attachment. You join a call, someone shares their screen, and that’s the whole attack surface.

The flaw lives in Zoom’s annotation feature, the one you use to scribble arrows on a colleague’s slide deck. A Security explained in a blog post that because Zoom’s client “automatically parses whatever it receives” while annotation is active, an attacker could send a “specially crafted message to corrupt the receiving client’s memory and run code on it.”

Here’s the part that makes it worse. The protocol inside the app opens a direct channel between the person viewing and the person sharing. So every participant on that call could be picked off individually.

End-to-end encryption didn’t save you

The vulnerability was apparently present in every version of Zoom, on every operating system. And it was exploitable even on calls with end-to-end encryption turned on, which is the setting people flip specifically because they assume it’s the safe one.

It’s been patched. But a patch only helps if you install it, and Zoom is exactly the kind of app people leave running for months without restarting. Go check your version.

The discovery method is the actual story

A Security says the models identified and exploited the issue in fewer than 20 prompts. The researchers steered them, providing direction on what to look for, so this isn’t a story about AI autonomously breaking software. Real expertise did the aiming.

But the aiming used to be the cheap part and the finding used to be the expensive part. That ratio just moved.

The frontier labs have been selling a different story

Every major AI lab has spent the past stretch talking up how capable their top-tier models are, and how that capability is precisely why access has to be gated. Too powerful for the wrong hands.

Meanwhile the tools already in everyone’s hands are reshaping the cyber industry anyway. That gap between the marketing and what’s shipping publicly is worth sitting with.

Last week at the Black Hat cybersecurity conference in Las Vegas, officials from the United States and the United Kingdom warned that the speed at which people are discovering vulnerabilities is quickly outpacing the ability to patch them.

What to do about it

Update Zoom. Today, not eventually. That’s the one concrete action available to you here, and it closes this specific hole.

The broader picture is less tidy. Safety was never guaranteed, but the number of people who can now find a zero-click RCE in widely deployed software just expanded well beyond the specialists, and most of us are more exposed than we realize.