If you shared your screen on a Zoom call any time before this week's patch, a silent attacker sitting in that same meeting could have owned your machine. No click required. No suspicious prompt. Nothing on your screen to suggest anything had happened at all.

The vulnerability lived in Zoom's annotation feature — the tool that lets participants draw arrows and type notes over a shared screen. Researchers found that the feature failed to properly validate input from other meeting participants, meaning a malicious attendee could send specially crafted annotation data to execute arbitrary code on the presenter's machine. The reverse was also true: a presenter sharing their screen could have compromised every participant watching. One bug, bidirectional blast radius.

What makes this particularly striking is how it was found. Researchers used a publicly available AI tool and reached the exploitable condition in fewer than 20 prompts. That's not a story about AI being clever — it's a story about how dramatically the economics of vulnerability research have shifted. What might have taken a skilled researcher days of manual code review now takes an afternoon with a chatbot and some persistence.

Zoom has patched the flaw, tracked across multiple CVEs, and says it has found no evidence of exploitation in the wild. If you're on Zoom desktop, check that your client has auto-updated. The affected versions span Windows, macOS, and Linux clients.

Why this matters beyond the patch. Zoom remains one of the most universally deployed pieces of software on the planet. It's running in hospitals, law firms, government departments, and every corporate boardroom that survived the pandemic. A zero-click, in-meeting RCE is about as bad as it gets for a collaboration tool — you don't even have to trick someone into joining a malicious meeting, you just have to be in a legitimate one with them.

The Australian context is worth noting. Zoom is approved for use across large portions of the Australian public sector, and many Australian enterprises rely on it for sensitive client communications. The ASD's Essential Eight requires application patching within 48 hours for internet-facing services with critical vulnerabilities — this one qualifies.

The AI-assisted discovery angle also deserves a longer look. OpenAI launched GPT-5.6-Cyber this week, a model explicitly tuned for vulnerability research and exploit development. The timing is awkward in the best possible way: a demonstration that commodity AI tools can find serious flaws lands the same day a lab ships a model purpose-built to do exactly that, faster and at scale.

The question defenders need to sit with isn't whether AI will be used to find bugs. It already is. The question is whether your patching cadence, your attack surface management, and your incident detection are moving at the same speed as the tools being pointed at you.

For now: patch Zoom, check your meeting security settings, and maybe don't assume that a familiar face in the participant list means a safe meeting.