‘Zoomsday’ hack uncovered using fewer than 20 AI prompts
Zoom has patched a major security vulnerability that could allow an attacker to hijack anyone's device during a meeting. In a blog post on Tuesday, researchers at A Security say they uncovered the flaw using "fewer than 20 prompts on publicly available AI models," as reported earlier by Wired.
The security research community has a new parlor trick, and it is not the kind that stays in the parlor. A flaw in Zoom's annotation feature, patched this week, could have let an attacker run code on any participant's device during a meeting. The discovery method is the real headline: fewer than 20 prompts on public AI models. That is not a brute-force assault or a zero-day bought on a dark web forum. It is a conversation.
The vulnerability itself is textbook in its mechanics. The annotation tool, designed for harmless screen drawing, became a vector for remote code execution. An attacker who could join or host a meeting could steal data, switch on a camera, or drop malware. No user action beyond attending the meeting. The attack surface was not exotic; it was a feature everyone trusted.
What matters here is the speed and cost of discovery. The researchers at A Security did not spend months reverse-engineering binaries. They asked an AI model a series of pointed questions, and the model connected the dots. This is the new reality of vulnerability research: the barrier to entry has dropped from a specialized skill set to a well-worded prompt.
For remote work, the implication is quiet but firm. Video conferencing is the backbone of distributed teams, and every feature added to that backbone is a potential lever. The annotation tool was a convenience, not a critical function. Yet it nearly became a backdoor into thousands of corporate networks. The lesson is not to abandon the tools, but to recognize that their complexity grows faster than their security.
The patch is out, and the window is closed. But the method is now public. Every attacker with a subscription to a large language model can replicate the approach, not just against Zoom but against any software with a similar feature set. The next vulnerability may not be announced in a blog post. It may simply be used.
This is the new arms race, and it is asymmetric. Defenders still need to audit code line by line. Attackers just need to ask the right questions. The quiet truth is that AI has not made hacking easier; it has made thinking faster. And in security, thinking faster is often the difference between a patch and a breach.