Zoom Hacked in a Single Day: AI Finds a Zero-Click Flaw in Annotations

Researchers at A Security have reported a series of vulnerabilities in Zoom that allowed a participant in a video meeting to execute arbitrary code on the devices of everyone else in the call. The attack required no action from the victim — no click, no file download, not even a visible sign of compromise. Simply being in the call was enough. The problem affected clients on Windows, macOS, iOS, Android and Linux across all versions up to and including 7.0.5.

Where the bug was hiding

The vulnerability lived in the annotation feature — the one that lets you draw, insert text and add shapes over a shared screen. Annotations are built on a proprietary protocol with no public documentation, and the client automatically parses any message it receives. The researchers found that an “acknowledgment” message (ack) could carry a full object description, and the recipient dutifully reconstructs it. A specially crafted packet triggered a buffer overflow: the length counter was taken straight off the wire, and there was no bounds check against the fixed 128-byte buffer. The result — a stack overwrite reaching all the way to the return address.

In the demonstration against macOS, an overflow through a single packet handed the attacker full control over the registers, and since the build had neither PAC nor a stack canary, it ended with Safari launching on the victim's machine. For Android, the researchers demonstrated an ASLR bypass via the heap layout and a partial overwrite of the pointer to the virtual function table — without leaking any addresses.

Three identifiers were assigned in total: CVE-2026-53413 (stack overwrite), CVE-2026-53414 (out-of-bounds read) and CVE-2026-53415 (an independent “write-anywhere” bug).

The real news is the speed

The authors of the report openly admit it: the mere existence of a hole in a popular product is secondary. Far more important is that the entire path from discovery to a working exploit was covered using publicly available AI models, in fewer than twenty prompts and in less than a day. Zoom is closed-source software with no published protocol, and work like this used to be considered the domain of state intelligence agencies: elite teams, months of effort, and budgets that governments regulate like weapons.

The AI agent first statically ranked 3,762 functions by risk, then dynamically traced which libraries load when specific meeting features are used, and homed in on libannotate.so — which the static analysis had ranked only 45^th.

What users should do

Zoom was notified of the vulnerability on June 10, 2026. The company shipped a client-side fix in version 7.1.0 twelve days later, and on July 15 it added server-side filtering of malicious messages for those who hadn't yet updated. The catch: the server-side filter doesn't work in calls with end-to-end encryption, where the server by definition can't see the contents. Updating to versions 7.1.5 and 7.0.6 closes all three CVEs; until then, owners of older clients are advised to disable end-to-end encryption, restrict access to meetings and trim optional features — each of them just one more parser of someone else's bytes.

Source: Lobsters