A Zoom Participant Could Reach Every Other Device. AI Built the Exploit in a Day.
Zoom has patched a set of memory-safety vulnerabilities that researchers combined into a zero-click remote-code-execution attack through the live-meeting annotation protocol. A malicious participant could target a presenter, while a malicious presenter could reach participants across Windows, macOS, iOS and Android.
The research team at A Security called the chain Zoomsday. It said an AI agent using public models produced a working exploit in one day with fewer than 20 prompts. Zoom fixed the issues before the researchers published technical details.
The meeting itself carried the attack
The attack surface was Zoom’s annotation protocol, which distributes drawing and presentation data among meeting participants. The target did not need to click a link, open a file or approve a prompt. Processing the malicious annotation traffic was enough to reach vulnerable native code.
A Security initially targeted the Android client and then confirmed that the vulnerable code path existed across desktop and mobile platforms. The team described two directions of attack: presenter to participant and participant to presenter.
The chain used three Zoom flaws:
- CVE-2026-53413, a high-severity buffer overwrite.
- CVE-2026-53414, a medium-severity buffer over-read.
- CVE-2026-53415, a high-severity use-after-free.
Zoom also published CVE-2026-53416, a separate high-severity path-traversal issue affecting its VDI client. It appeared in the same bulletin period but is not presented by A Security as part of the annotation exploit chain.
Why the zero-click condition changes urgency
Meeting software is designed to accept complex real-time input from people outside an organisation. That makes a live call a temporary trust relationship with every participant. The annotation feature crossed that relationship into native memory handling without a separate consent step.
The researchers said the attack produced no useful visual warning for the target. A host might therefore believe a call was routine while malicious protocol messages were being processed in the background.
A Security demonstrated code execution in a controlled setting. Neither the researchers nor Zoom reported confirmed malicious exploitation in the wild when the findings were disclosed.
AI shortened the exploit-development cycle
The AI component is consequential because it was used for exploit engineering, not merely vulnerability summarisation. According to the researchers, the agent navigated the target, developed the primitives and produced a working chain with limited prompting.
That does not prove that any attacker can automatically weaponise every memory-corruption flaw. The team supplied expert direction and a suitable target. It does show that the time between technical disclosure and practical exploitation can compress sharply once enough detail is available.
Updates and mitigations
Zoom released fixed builds across Workplace, Rooms, the Meeting SDK and VDI products. Cloud Link’s consolidation of the vendor guidance lists fixed Workplace releases including 7.1.5 and 7.0.6, with product-specific builds for the other affected lines.
- Update every client platform. Do not assume that patching Windows covers phones, meeting rooms or embedded SDK deployments.
- Inventory Meeting SDK use. Applications that embed Zoom may require their own rebuild and release.
- Restrict untrusted meetings until updated. Meeting passwords and waiting rooms reduce unwanted attendance but do not neutralise a malicious authorised participant.
- Monitor crashes and child processes. Unexpected Zoom client termination or process creation can be useful investigation signals.
- Accelerate patch testing. Public AI-assisted exploit development reduces the safety margin after technical publication.
The BlackTree view
Zoomsday is a reminder that collaboration software is an internet-facing parser with microphones and cameras attached. Trusting someone enough to admit them to a meeting should not mean trusting every byte their client can send to native code.
The defensive window is also changing. When researchers can turn a complex cross-platform memory-safety finding into working code in a day, organisations need patch processes measured against exploit-development time, not against the next monthly maintenance window.
Sources
- A Security, Zoomsday: Zero-Click Remote Code Execution in Zoom via the Annotation Protocol, published 11 August 2026. The page provides no publication time.
- Zoom Security Bulletin index, vulnerabilities published 11 August 2026 and updated 14 August 2026. The index provides no publication times.
- Cloud Link, Zoom patches Zoomsday zero-click RCE, published 11 August 2026 at 16:04 CET.
