Zoom 'Zoomsday' Flaw Exploited in Under 20 AI Prompts

A critical vulnerability in Zoom's annotation feature — dubbed "Zoomsday" — was patched on Tuesday after researchers at A Security demonstrated they could develop a working exploit using fewer than 20 prompts on publicly available AI models, completing the work in a single day. The flaw affected Zoom across all five major platforms: Windows, macOS, Linux, Android, and iOS, and required zero interaction from victims to execute.

The practical impact was severe. An attacker needed only to join or host a meeting and trigger the annotation feature — the tool that lets participants draw over a shared screen — to run arbitrary malicious code on every other participant's device. From that foothold, the attacker could steal data, silently activate cameras or microphones, or install malware. A Security noted that the compromise left "no visual cue indicating the compromise," meaning victims had no in-session signal that anything had gone wrong.

The choice of attack surface is significant. Screen-sharing annotation is a broadly trusted, UI-level feature that rarely receives the same scrutiny as network-layer code or authentication flows. Vulnerabilities embedded in collaborative interaction features are particularly dangerous because they activate during high-trust sessions — exactly when a victim is least likely to notice anomalous device behaviour. This is consistent with a broader pattern of AI-assisted tooling exposing attack paths embedded in workflows users treat as inherently safe.

The headline figure — fewer than 20 AI prompts — is deliberately chosen by A Security to benchmark offensive research effort. Vulnerability researcher Idan Levcovich framed the historical baseline explicitly: producing a working exploit against this class of flaw "has always been nation-state work: elite teams, months of effort, budgets that governments regulate as weapons." A Security collapsed that to a single day, using an AI agent and models available to any practitioner.

That compression of effort has a direct implication for threat modelling: the cost curve for discovering and operationalising critical vulnerabilities is dropping, and the drop is now measurable in prompt counts rather than person-months. AI bug hunters are already appearing on Linux security mailing lists, and the Zoom case confirms that trajectory is moving beyond research novelty into reproducible practice.

The source material does not enumerate the specific prompts used, but the structure is clear: A Security deployed an AI agent — not a human analyst manually querying a chatbot — against the target. The agent-mediated approach implies iterative, automated reasoning over the codebase or protocol behaviour rather than one-shot code review. Fewer than 20 prompts to a working exploit suggests either a highly focused attack hypothesis at the outset, or an agent capable of collapsing multiple investigative steps into single high-density queries — likely both. For security teams evaluating where AI fits in their own pipelines, this is a concrete data point on what the agentic era means operationally.

Zoom issued the patch on the same day as the disclosure, indicating coordinated responsible disclosure. The broader signal, however, is not about any single CVE. When publicly accessible AI models can compress nation-state-grade offensive research into a sub-20-prompt, single-day workflow, the assumption that complex vulnerabilities carry implicit obscurity protection needs immediate recalibration — and so does any defender's research budget that hasn't yet accounted for it.