One Researcher, Under 20 Prompts, One Day: How AI Built a Zoom Attack
A Security disclosed three critical Zoom flaws on August 11. The bug is bad enough, but the finding is the cost: work that used to take nation states months took a single researcher a day.
On August 11 the Israeli security firm A Security published research it calls Zoomsday: three critical vulnerabilities in Zoom that let anyone sitting in a meeting take over another participant’s device. No click, no download, nothing required of the victim beyond being in the call, which makes it a zero-click attack: the kind where there is no user mistake to point at afterwards. The flaws are tracked as CVE-2026-53413, CVE-2026-53414 and CVE-2026-53415, all rated 9.0 on the critical scale, and they affected Zoom on Windows, macOS, iOS, Android and Linux. Zoom has patched them.
The bug lives in Zoom’s annotation feature, the tool that lets people draw and add text on a shared screen. Annotations do not travel as pictures. Each mark becomes a small structured message that the receiving client unpacks and rebuilds. The receiving code trusted a size number supplied by the sender, so a specially crafted message could write past the end of a buffer and, from there, take control. The researchers demonstrated it live by making a victim’s Mac open Safari from a single message. Because annotation is always on, the feature did not have to be in use for the attack to work.
The headline finding is not the bug. It is the price. A Security found the flaw and built a working exploit in under 24 hours using fewer than 20 prompts on publicly available AI models. Zoom is closed source with no published protocol, exactly the sort of target that used to require an elite team, months of reverse engineering and a budget governments regulate as a weapons export. The researchers’ own framing is blunt: that barrier has collapsed and it is not coming back.
Two things are worth holding at once here. The same capability that makes attacks cheap also makes defence cheap, and this is the third such story in a fortnight, after OpenAI’s GPT-5.6-Cyber found two Chrome zero days and PortSwigger’s AI system uncovered novel HTTP attacks. AI is not making anyone an instant hacker; skilled researchers still had to know which surface was worth attacking and when the model was wrong. What has changed is that experimentation now costs almost nothing, so the gap between “a bug exists” and “somebody has a working exploit” is shrinking on both sides. The uncomfortable corollary is that a published patch is a map to the bug, so every day your software stays unpatched is worth more to an attacker than it used to be.
What this means for you: update Zoom now, to version 7.1.5 or later. That is the whole action item and it takes a minute. If you administer Zoom for a company, you can set a minimum client version for guests as well as staff, and turning off features nobody uses, such as annotation, whiteboarding, remote control and file transfer, removes whole categories of future bug. One counterintuitive note from the disclosure: end to end encryption actually left older clients more exposed here, because Zoom’s server side filter cannot inspect traffic it cannot read. More broadly, the lesson for everyone is unglamorous. Patch quickly, and treat “we are too obscure to be attacked” as a retired argument.
Sources
That Advice About Which Language AI Codes Best In? It Falls Apart on Real Work
Dan Luu re-ran the popular claim that dynamic languages are cheaper for coding agents. On two large tasks the effect vanished, and the second-most-cited benchmark turned out to be broken.
