Zoom patched three vulnerabilities in its meeting-annotation feature that could have let a participant in a call remotely execute code on another participant’s device.
The research group at A Security dubbed the vulnerability “zoomsday” and said that, if exploited, the bug would allow an attacker to remotely take over and execute code on the device of another participant in the call.
The flaw, researchers said, targets attendees individually rather than the whole meeting at once. The most severe of the three flaws, tracked as CVE-2026-53413, was present in Zoom Workplace clients on all supported platforms before version 7.1.5 or 7.0.6, depending on release branch, as well as in the Windows VDI Client, Zoom Rooms and the Zoom Meeting SDK before their respective fixed versions.
Worse, A Security says the entire attack, from finding the flaw to building a working exploit, took its researchers fewer than 20 prompts (the written instructions researchers fed to publicly available AI models, directing them to do the work) in under 24 hours.
To Click or Not to Click: That is the Question
The vulnerability was issued three CVE classifications, CVE-2026-53413, CVE-2026-53414, and CVE-2026-53415. Zoom’s bulletins score CVE-2026-53413 and CVE-2026-53415 at 8.3 and CVE-2026-53414 at 6.5 under CVSS 3.x. A Security’s own writeup puts all three at 9.0 under CVSS 4.0.
A Security said the exploitation required nothing from the victim beyond being in the call: no click, no download, no warning.
Zoom’s advisory describes things slightly different. In Zoom’s account of the three vulnerabilities (CVE-2026-53413, CVE-2026-53414, and CVE-2026-53415) in its posted CVSS score, it includes a “user interaction” field indicating whether a victim has to do something such as click a link, open a file, in order for the attack to work. Zoom’s bulletins mark that field “required” for all three CVEs.
That contradicts A Security’s account, where the only thing asked of the victim is being in the call; annotation runs in the background and updates automatically, so nothing on the receiving end needs to be clicked or approved.
Typically, a CVSS field helps security teams decide what gets patched first. A flaw marked “user interaction required” can get triaged behind one marked “no interaction needed,” even when, as here, the practical risk looks the same either way.
Mismatches like this aren’t unusual across the industry, in general. A CVE’s severity score is typically assigned by whoever owns the tracking process, often the vendor itself. This is based on how that company understands the flaw at the time of scoring. Outside researchers who find the same bug are scoring it against their own exploit, which may use a different path to trigger it or rest on different assumptions about what a “typical” attack looks like.
Neither score is necessarily wrong. They’re often just answering “how easy is this to exploit?” from two different vantage points – one internal, one adversarial.
In this case, the gap runs deeper than that general pattern. Zoom scored the three CVEs under CVSS 3.x. A Security’s own writeup scores them under the newer CVSS 4.0 framework, which weighs some factors differently. So the two numbers aren’t strictly measuring the same scale.
And on the “user interaction” question specifically, the two accounts aren’t just weighing the same facts differently. They describe different requirements for the attack to work.
AI Prompting and Probing for Bugs
The bug hunting itself was conducted in part using AI and reverse engineering the Zoom code. A Security’s researchers first had an AI model scan Zoom’s Android app and rank thousands of internal functions by how likely each was to contain a memory-handling bug. This is the kind of manual triage that normally eats days of a human researcher’s time.
That first pass pointed the wrong direction, so the researchers redirected the model to trace how Zoom’s screen-annotation feature – the tool that lets someone draw or type on a shared screen – builds and sends its data between participants during a call.
That redirection led to the actual flaw which was a single field in the annotation code that trusted a number sent over the network to decide how much data to copy into a fixed-size memory buffer, with nothing checking that the number wasn’t larger than the buffer itself.
From there, the AI modeled a step-by-step exploit chain. First, it was on Mac clients, a technique that hijacked a corrupted function return to launch an unrelated app on the victim’s machine — to prove the bug was more than theoretical.
Using AI to find bugs is relatively new, but quickly becoming standard. HackerOne reported in 2025 a 210% year-over-year jump in valid AI-assisted vulnerability submissions.
According to the research team, the flaw exists in the way Zoom handles annotation during meeting sessions. Because those requests do not get properly checked for sanitized requests, the attacker can insert commands that will be instantly run by the victim.
“Every Zoom client automatically parses whatever it receives, sending a specially crafted message to corrupt the receiving client’s memory and run code on it,” A Security explained. “The proprietary protocol used by the annotator opens a direct channel between a viewer and a sharer, thereby targeting each participant individually.”
Disclosure Timeline
The team at A Security contacted Zoom and thee flaws were patched before the exploits were exploited in the wild, according to A Security. Researchers said it discovered the flaws June 8 while analyzing Zoom’s Android client, confirmed a working exploit against Zoom 7.0.5 the following day, and privately reported the issue to Zoom on June 10.
Zoom shipped an initial client-side fix in version 7.1.0 on June 22, added a server-side mitigation on July 15, and issued a further client fix in version 7.1.5 on July 20 covering the remaining flaw.
Not every install received that fix automatically. Zoom’s standard desktop client updates itself by default, but enterprise deployments using Zoom’s MSI or PKG installers only auto-update if an administrator has specifically configured it to. This is common in corporate environments. Mobile clients depend on each user’s own app-store update settings.
Organizations running Zoom at scale should confirm their deployment is actually on a patched version rather than assume the fix has already reached every device.
As is proper security practice, the details of the bug are only now being released as devices have had several weeks to receive secured updates from Zoom.
Researchers warn AI assisted bug hunting has dramatically lowered the bar for adversaries to find and exploit similar vulnerabilities.
“Acquiring one has always required nation-state infrastructure, elite teams, and months of work. That barrier is gone,” A Security wrote in its post on the research. “[A Security] found a critical, nation-state-grade vulnerability in Zoom in a single working day using an AI agent and models anyone can access today.”
Image by Mohamed Hassan from Pixabay
