Critical ‘Zoomsday’ flaw enables total device takeover during Zoom calls — AI-assisted research only used 20 prompts to find an exploit to hack hundreds of millions of people.

We’ve typed many words about how the industry-standard 90-day security bug disclosure window is effectively dead and gone with the advent of AI-assisted exploiting. Illustrating that point rather poignantly, researchers at A.Security easily came up with Zoomsday. This exploit let any participant in a Zoom meeting gain control over the device of anyone else, all without them being any wiser.

The team claims it cooked the exploit with merely 20 prompts to an AI agent. The exploitable area is substantial, as recent estimates pin Zoom’s monthly active users at around 220 million and an estimated 56% of the global conferencing market share.

There were two remote code execution (RCE) vulnerabilities present in Zoom Workplace before 7.0.6 and, for users on the “fast track” branch, before version 7.1.5. The bugs were in a library used by Zoom’s annotation functionality, though no participants need to actually use the whiteboard for the exploit to work — the code is always on, so all an attacker needed to do was join the meeting. Zoom quickly fixed the bugs after the initial reports, so everyone who has updated Zoom Workplace to the current version should be safe.

With the exploit, the attacker was able to get full remote code execution, meaning they could effectively control the user’s computer and their data — invisibly, to boot. Zoom isn’t an application that runs with administrator privileges, so kernel-level rootkits are off the menu, but once you have the user’s data, it’s not like you need much else. Plus, it’s easy to gain exploit persistence any number of other ways.

A.Security says a small team developed this exploit with a mere 20 prompts to an AI agent — pointing out how easy it was to come up with a nation-state-class vulnerability with meager resources. While the majority of AI-assisted vulnerability research focuses on open-source software or applications with published communications protocols or file formats, Zoom is fully proprietary, and it was still easily cracked open.

The firm further noted that “the model requiring elite teams, months of effort, and weapons-grade budgets has collapsed,” and that “the barrier that kept these weapons scarce has collapsed, and it will not come back” — basically repeating what every security researcher has been yelling from the top of their lungs for the past year or so.

The vulnerability itself is, rather unsurprisingly, a buffer overrun: the program fails to check that an input is the right size, so you can push more data than it expects and overwrite part of the following memory with code that will be executed.

First, the scientists decompiled the Android package and asked an AI agent to rank the potential attack surfaces to relatively little success. They then turned their attention to the communications protocol. They found that the code library handling annotations received each object (rectangles, text, etc.) in serialized form, with count fields telling the recipient how much data to read next.

Crucially, they found that the code handling these reads didn’t have a boundary check for maximum size, meaning one could simply lie about it and send a chunk of data that’s too large and padded with exploit code at the end, as Norman Stansfield would say, bin-go!