A team at A.Security pointed an AI agent at Zoom, fewer than twenty prompts and less than a day later, they had a working exploit that would let anyone in a meeting silently take over everyone else’s computer.

The flaw they named Zoomsday, catalogued as CVE-2026-53413, is a zero-click remote code execution bug in Zoom Workplace before version 7.0.6, or before 7.1.5 on the fast-track branch. It lives in a library used by the annotation feature, the whiteboard tool that lets participants draw on a shared screen. Nobody has to open the whiteboard for the attack to work. The code is always running, so joining the meeting is enough. Zoom patched it after the disclosure. For scale, recent estimates put Zoom at around 220 million monthly active users and roughly 56% of the global conferencing market.

A buffer overrun, about the oldest mistake in software. The annotation library receives drawing objects, rectangles, text and so on, in serialised form, each one preceded by a count field announcing how much data follows. The code read that count and believed it. Send a count that lies, follow it with more data than the buffer was sized for, and the extra data spills into the memory sitting next door. Pad the end of that extra data with your own instructions and the program runs them.

The result is full remote code execution, invisible to the person it happens to. Zoom does not run with administrator privileges, so a kernel-level rootkit is off the table. That is less comfort than it sounds. By that point the attacker already has the user’s files and session, which is usually what they came for.

The researchers first decompiled the Android package and asked the agent to rank potential attack surfaces, which worked poorly. They then turned it on the communications protocol instead, and that is where the missing bounds check surfaced. The detail worth holding on to is that Zoom is fully proprietary. There is no published protocol, no source code, no file format specification. Most AI-assisted vulnerability research so far has gone after open-source projects or documented formats, precisely because those are easy to read. Zoom is neither, and it still took a day.

A.Security’s own framing is blunt: the model requiring elite teams, months of effort and weapons-grade budgets, they say, has collapsed. Estimates quoted alongside the disclosure put the conventional cost of a find like this at about five people working for six months.

The security industry’s ninety-day disclosure convention rests on a rough balance: finding a serious bug takes about as long as fixing one, so a researcher can afford to wait. If discovery drops to a day while patching, testing and enterprise rollout stay at weeks, that convention stops describing anything real. Nothing about Zoomsday is clever. A missing bounds check is a first-year error, and it had presumably been sitting there for years.

Sources: Tom’s Hardware; A.Security disclosure; Security Boulevard; CVE-2026-53413.