In the world of cybersecurity, the narrative has long been dominated by the specter of powerful, frontier artificial intelligence (AI) models. Those AI models cloistered behind restricted access, touted as too dangerous to be freely available. Yet, a recent discovery has upended that assumption, revealing that even publicly accessible AI tools, when guided by skilled hands, can expose vulnerabilities capable of shaking the foundations of digital trust.

The case in point is Zoom, the ubiquitous video conferencing platform that became a lifeline for businesses and individuals alike. A cybersecurity firm uncovered a flaw so severe it bordered on catastrophic: a zero‑click remote code execution vulnerability embedded within Zoom’s annotation feature. The exploit allowed attackers to hijack devices simply by being present in a call with screen sharing enabled. No clicks, no downloads, no suspicious links, just the silent corruption of memory through a specially crafted message.

What makes this incident remarkable is not just the breadth of exposure—every version of Zoom across all operating systems was affected, even those with end‑to‑end encryption—but the method of discovery. The firm reported that the vulnerability was identified using publicly available AI models, requiring fewer than 20 prompts and less than 24 hours to surface. This was not the work of cutting‑edge, restricted systems but of accessible tools, sharpened by the expertise of human researchers who knew where to direct their focus.

The implications are profound. For years, policymakers and AI labs have emphasized the risks of frontier models, warning that their immense capabilities could be weaponized if they fell into the wrong hands. The speed at which vulnerabilities are being discovered now outpaces the ability of companies to patch them, a reality underscored by warnings from U.S. and U.K. officials at the Black Hat conference in Las Vegas.