Zoom has shipped fixes for three related vulnerabilities in its Annotation tool — the feature that lets meeting participants draw or leave text notes over a shared screen. Researchers at A Security, who found the bugs, say they went from analyzing Zoom’s proprietary protocol to a working proof-of-concept in under 24 hours, using fewer than 20 prompts to publicly available AI models.
What was actually broken
The Annotation feature doesn’t send drawings and text as flat images. Instead, it serializes them into structured objects that travel through Zoom’s infrastructure and get rebuilt on the receiving end. A Security found that the code responsible for reconstructing those objects didn’t validate incoming data carefully enough, opening the door to memory-corruption bugs.
The most serious of the three, tracked as CVE-2026-53413, is a buffer overflow that Zoom itself rates 8.3 on the CVSS scale — a “High” severity score, not “Critical.” The second, CVE-2026-53414 (6.5), is a buffer over-read that can crash another participant’s client rather than run code on it. The third, CVE-2026-53415, is a use-after-free flaw also rated 8.3.
The “zero-click” claim doesn’t quite match Zoom’s own numbers
A Security branded its find “Zoomsday” and described it as an attack requiring nothing from the victim beyond simply being in the call — no click, no download, no warning. That framing is what most coverage ran with.
Worth flaggingZoom’s own published CVSS vectors for all three bugs mark user interaction as required, not absent — which sits awkwardly next to a “zero-click” label. The underlying bug is real and the severity is genuinely high, but how invisible an actual attack would be depends on whose account you take at face value: the vendor’s technical scoring, or the researcher’s marketing framing.
What a successful exploit could theoretically do
If pulled off, the bug could let an attacker run arbitrary code on a target’s machine — enough, in principle, to reach files, exfiltrate data, drop further malware, or maintain a foothold on the system. Depending on what access was gained, camera and microphone exposure were also on the table.
None of this has been reported as an active, in-the-wild campaign. What exists is a research demonstration against test devices, reported privately to Zoom and patched before A Security published its findings — not a confirmed attack against real users.
The AI part, minus the hype
This is where the story gets genuinely interesting for the security industry, and also where it’s easiest to overstate what happened. A Security’s own writeup leans toward crediting the discovery to “publicly available frontier models.” More detailed technical accounts describe something narrower: a human researcher manually reverse-engineered Zoom’s undocumented protocol, and used AI models along the way to speed up specific steps — ranking which native functions were worth digging into, and spotting unsafe memory operations faster than manual review alone.
In other words, the model accelerated the work rather than autonomously finding and weaponizing the bug end to end. The headline number — under 20 prompts, under a day — is accurate and still notable. What used to take a well-resourced team weeks or months of grinding through undocumented binary protocols took a fraction of that time here. That cuts both ways: the same tools available to defenders and researchers are available to anyone else running the same prompts.
Zoom has already fixed it
All three annotation bugs are patched in current Zoom releases. As with any disclosure like this, the practical advice is unglamorous but real: update the client, don’t wait for a reminder prompt, and treat this as more urgent on machines running older builds in enterprise environments where patching lags.
This wasn’t the summer’s only serious Zoom bug
The annotation flaws aren’t connected to another vulnerability Zoom patched a month earlier. CVE-2026-53412, disclosed in July, carries a CVSS score of 9.8 — genuinely critical — and could have let an unauthenticated attacker take over a Zoom account over the network, on Windows clients, without any action from the victim. One detail worth correcting from early coverage: Zoom’s revised bulletin, published July 15, removed the Meeting SDK from the list of affected products — so despite what some initial reports said, that particular component is no longer implicated.
Taken together, Zoom spent the summer closing two very different kinds of holes: a critical, unauthenticated account-takeover bug in July, and a set of high-severity annotation flaws in August that came with an unusually fast, AI-assisted research timeline attached.
The takeaway
- Update Zoom now — this applies more than usual to managed, enterprise-deployed installs.
- There’s no evidence of a mass exploitation campaign; this is a patched research finding, not a confirmed breach.
- Treat “zero-click” claims from vendor-adjacent security research with a bit of caution when the vendor’s own CVSS scoring says otherwise.
- The real signal here isn’t one Zoom bug — it’s how much AI tooling is compressing the time between finding a flaw and having something that works against it.
Sources: SecurityWeek · Security Affairs · The Hacker News · CSO Online · BleepingComputer · A Security (Zoomsday disclosure) · Zoom Security Bulletins ZSB-26014, ZSB-26015, ZSB-26016, ZSB-26017
Більше новин в телеграмі
