# Researchers used AI to build a Zoom hijacking exploit in one day > Source: > Published: 2026-08-12 00:36:44+00:00 *A security researcher fed AI models fewer than 20 prompts and had a working exploit for a critical Zoom bug in under 24 hours, and Zoom just confirmed the flaw let anyone in a meeting silently take over another attendee's device.* Zoom published a security bulletin on August 11 disclosing three vulnerabilities in the screen-share annotation tool, the pen and highlighter feature you've probably ignored in a hundred meetings. The worst of them, CVE-2026-53413, carries a CVSS score of 8.3. It lets a meeting participant run code on someone else's computer with no click, no download, and no warning popup. Researchers at A Security have taken to calling it "Zoomsday." ## How the Zoomsday bug works The bug lives in a routine called CAnnoFormatBlock::Deserialize, buried inside Zoom's annotation library, libannotate.so. It's an old kind of flaw. Software has carried it for decades: a fixed 128-byte buffer that trusts a 32-bit character count arriving over the network instead of checking it first. Send an oversized value and the parser copies more data than the buffer can hold. It spills into adjacent memory. No click required. In a proof-of-concept, that overflow was enough to launch Safari on a victim's Mac without them touching anything. Idan Levcovich, a senior vulnerability researcher at A Security, found the bug on June 8 while picking apart Zoom's Android client. He had a confirmed working exploit against Zoom 7.0.5 the next day. "One malformed annotation, drawn by anyone in the call, was enough to take over every other device in the room," Levcovich said, according to Security Boulevard. A Security reported the flaw to Zoom on June 10. Zoom moved fast - at least by enterprise software standards. Client-side fixes started rolling out June 22, a server-side mitigation followed on July 15, and the public bulletin landed August 11, alongside two related bugs: CVE-2026-53414, a lower-severity buffer over-read that can crash a client, and CVE-2026-53415, another annotation-parsing flaw involving attacker-controlled pointers, also rated 8.3. The fixes cover Zoom Workplace, the Windows VDI Client, Zoom Rooms, and the Meeting SDK, spanning Windows, macOS, Linux, iOS, and Android. Fixed versions include Workplace 7.1.5 (or 7.0.6 on the legacy end-to-end encrypted branch) and VDI Client 7.0.11 or 6.6.16. ## AI shrank the timeline to a day Buffer overflows aren't new. What's new is the speed. Levcovich has said the reverse-engineering, the protocol reconstruction, and the hunt for unsafe operations in the annotation code leaned heavily on publicly available AI models, the kind anyone can sign up for today. Building a reliable, cross-platform, zero-click RCE chain against a Fortune 500 communications platform used to be the sort of project that took a well-funded team months, the domain of nation-state hacking units with real budgets. A Security says this one came together in a day, from a handful of prompts. That's the part enterprise security teams should sit with. Patch timelines, threat models, and vulnerability disclosure windows were all built around the old math: discovery is slow, weaponization is slower, and defenders usually have weeks before a proof-of-concept turns into something dangerous in the wild. If a single researcher can go from first look to working exploit in under 24 hours using off-the-shelf AI tools, that buffer is gone. Zoom had roughly two weeks between disclosure and its first client patch. Under the old timeline, that would've felt comfortable. Under this one, it's tight. None of this means Zoom handled it badly. A private report, a patch within days, a staged rollout with a server-side mitigation as backstop, that's close to the textbook response. The real exposure sits with the millions of businesses that run Zoom calls with screen sharing on by default and haven't pushed the update. Annotation is enabled out of the box in most Workplace deployments, and a lot of IT teams treat client updates as a background chore rather than a fire drill. Frankly, the lesson isn't really about Zoom. Every major platform with a rich-media feature, screen annotation, file preview, live transcription, has some version of this same attack surface: a parser trusting data it shouldn't. AI just made the search for that trust gap dramatically cheaper. Security teams that still budget vulnerability response in weeks are planning for a world that no longer exists. **Also read:** [CoreWeave's Q2 Revenue More Than Doubled to $2.58 Billion on AI Demand](https://startupfortune.com/coreweaves-q2-revenue-more-than-doubled-to-258-billion-on-ai-demand/) • [Brad Lightcap Is the Latest Senior OpenAI Executive to Head for the Exit](https://startupfortune.com/brad-lightcap-is-the-latest-senior-openai-executive-to-head-for-the-exit/) • [SpaceXAI and Cursor Launch Grok Bot, an Always-On Team of AI Agents](https://startupfortune.com/spacexai-and-cursor-launch-grok-bot-an-always-on-team-of-ai-agents/)