Zoomsday: Zoom zero-click RCE, AI wrote the exploit

Researchers at A Security disclosed a zero-click remote code execution chain in Zoom's screen-share annotation feature on August 11. Join a meeting, send a malformed annotation packet, run code on every other participant's machine. No click, no download, no dialog. The detail that should change how you plan patch windows: they built the working exploit in under 24 hours using fewer than 20 prompts against publicly available AI models, as TechRepublic reported.

What actually happened

The chain — nicknamed ZOOMSDAY — lives in how Zoom's native clients deserialize annotation formatting data. A missing bounds check means attacker-controlled length fields get trusted when data is copied into fixed-size buffers. Zoom's annotation protocol opens a direct channel between a screen viewer and a sharer, and every client parses whatever arrives automatically. That is the whole attack surface.

Zoom tracks three bugs from the chain, per Security Affairs:

• CVE-2026-53413 — stack buffer overflow in CAnnoFormatBlock::Deserialize, the RCE primitive
• CVE-2026-53414 — buffer over-read leading to denial of service
• CVE-2026-53415 — use-after-free

Researchers confirmed zero-click RCE against Zoom Client v7.0.5 on Windows, macOS, iOS, and Android. Fixes are in Workplace 7.1.5 and 7.0.6, Rooms 7.1.5, and Meeting SDK 7.1.5. We have seen no report of in-the-wild exploitation, and we are not going to imply one.

Why an AI-written exploit matters for your business

Ignore the CVE numbers for a second. The finding is the clock.

Your patch cadence was built on an assumption: after a bug is disclosed, someone still has to do days or weeks of skilled reverse-engineering to weaponize it. That gap was your grace period, whether you named it or not. Fewer than 20 prompts and one day is not a grace period. It is the same afternoon.

Two changes follow, and neither needs a security team:

Turn on auto-update for the software that parses untrusted input. Video clients, browsers, PDF readers, chat apps. These things ingest data from strangers by design. If your team updates Zoom manually "when they get a prompt," half your fleet is a version behind right now. Check yours today: Zoom 7.0.5 or older is the vulnerable line.

Stop treating feature surface as free. Annotation is a nice-to-have that shipped a memory-corruption bug reachable by any meeting participant. Every parser you enable is a parser someone can feed. If your org does not use in-meeting annotation, disable it at the account level — that is a policy toggle, not an engineering project. We made the same argument about browser agents touching your local machine: default-on convenience features are where the exposure actually lives.

The defenders now have the same models the researchers used. That cuts both ways, and it is the reason OpenAI and Google are shipping security-specific models. But tooling parity does not help you if your update policy still assumes weeks.

Sources: Security Affairs, TechRepublic.