In early June, researchers at the digital defense firm A Security turned a new generation of publicly available artificial intelligence models loose on one of the most widely used enterprise and consumer software platforms in the world—Zoom. The AI required fewer than 20 prompts to discover a set of critical vulnerabilities in the video conferencing application’s screen-sharing annotation system. On Tuesday, the company confirmed the flaws and began rolling out fixes, but the disclosure already raises uncomfortable questions about how quickly the landscape of cybersecurity threats is being reshaped by the very tools meant to defend it.
The vulnerabilities, which affect Zoom on Windows, macOS, Linux, iOS, and Android, could have allowed an attacker to silently hijack the device of any participant on a call where screen sharing was enabled. No interaction from the victim was required. No alert or suspicious prompt would have appeared. The attack surface was vast: anyone who joined a meeting and saw a shared screen—whether as host or participant—was potentially exposed. The mechanism lay not in Zoom’s core video or audio protocols, but in a feature that many users take for granted: the real-time annotation tools that let participants draw, highlight, or type on a shared screen during a presentation.
The Annotation Flaw That Opened Every Screen Sharer to Attack
Zoom’s annotation feature is a convenience that makes collaborative meetings more interactive. It is also, as A Security’s AI-driven auditing process revealed, a complex and opaque subsystem within a proprietary codebase—exactly the kind of corner where human reviewers often miss subtle but exploitable bugs. The researchers discovered that the protocol governing how annotations are transmitted and rendered contained multiple vulnerabilities that, when chained together, gave an attacker full control over a target’s device.
Because annotation data is sent in real time from one participant to all others, a malicious actor who was already on a call could craft specially malformed annotation packets. The receiving clients would process these packets without proper validation, leading to memory corruption, arbitrary code execution, and ultimately device takeover. The attack leveraged no user error—no phishing link, no download, no credential theft. It was a silent, interaction-free exploit that could be launched from within the meeting itself.
The implications for enterprise security are severe. Zoom has become a backbone of remote work, online education, healthcare consultations, and government communications. Screen sharing is one of its most common uses, and annotation is frequently enabled by default in many organizational settings. An attacker who managed to join a meeting—or who was a legitimate participant with malicious intent—could have compromised every other attendee without leaving a trace until the damage was already done.
How AI Discovered a Critical Vulnerability in Under 20 Prompts
What makes this disclosure particularly significant is not just the severity of the bug, but the method by which it was found. A Security used publicly available AI models—the same kind of large language models and code-generation tools that have been in the spotlight for their potential to accelerate both defensive and offensive cybersecurity work. The researchers directed the AI to analyze Zoom’s annotation protocol, a subsystem that, by their own account, they chose because AI models trained on common vulnerability patterns have learned to expect flaws in convoluted, proprietary, and seldom-reviewed features.
Omer Gull, co-founder of A Security, noted that before AI tools reached their current level of capability, such a discovery would have required a team of five people approximately six months of iterative refinement, reverse engineering, and manual testing. Now, with fewer than two dozen prompts—each essentially a natural language instruction to the AI to examine a specific component or attack vector—the team obtained a working exploit chain. The AI did not merely identify potential weaknesses; it proposed specific packet structures, validated them against the Zoom client, and iterated toward a fully functional proof of concept.
This democratization of vulnerability research is double-edged. On the one hand, defenders can now audit software faster and more thoroughly than ever before. On the other hand, the same tools are available to malicious actors. The barrier to entry for discovering serious, weaponizable bugs has dropped dramatically. Zoom, with its massive user base and high trust quotient, represents a prime target for anyone—state-sponsored groups, cybercriminal syndicates, or lone hackers—who wants to compromise systems at scale.
What Is the Zoom Device Hijack Vulnerability and How Does It Work?
To understand the technical nature of the flaw, it helps to know that Zoom’s annotation feature relies on a custom protocol for transmitting drawing commands, mouse movements, text inputs, and shape overlays between participants. This protocol must handle a variety of data in real time, with low latency, and across multiple platforms and devices. Because Zoom did not open-source this protocol or subject it to the same degree of public scrutiny as, say, an IETF standard, internal implementation errors went undetected.
A Security’s AI identified that the annotation packet parser lacked rigorous bounds checking and input sanitization. By sending a sequence of crafted packets, an attacker could overwrite critical memory regions in the Zoom client process. From there, the attacker could hijack execution flow, install persistent backdoors, exfiltrate data, or pivot to other systems on the same network. Because the attack happened entirely within the context of the Zoom application, antivirus and endpoint detection systems would have seen only normal Zoom traffic.
The vulnerabilities were present across all operating systems, indicating that the issue lay in Zoom’s cross-platform codebase rather than in any platform-specific library. This homogeneity of attack surface is a double threat: a single exploit works against a Windows machine, a MacBook, an Android phone, and an iPhone alike, multiplying the potential impact.
The Speed of Disclosure: From AI Discovery to Vendor Fix
A Security reported the vulnerabilities to Zoom in early June. Within weeks, Zoom had developed and began deploying patches. The company issued a security advisory on Tuesday detailing the flaws—cataloged as ZSB-26015—and urged users to update to the latest version. While the speed of the response is commendable, the incident underscores a growing reality: the window between vulnerability discovery and exploitation is shrinking. When an AI can find a critical bug in a few hours, attackers—who are not bound by responsible disclosure timelines—can do the same. The industry must adapt to a pace of vulnerability identification that outruns traditional patch cycles.
Zoom has not disclosed whether it observed any exploitation of these flaws in the wild, and as of this writing, no public reports of attacks have emerged. However, the passive nature of the exploit—no victim interaction required—makes it especially difficult to detect retroactively. Organizations that rely on Zoom should prioritize updating their clients immediately, as well as reviewing whether screen sharing and annotation features can be restricted or disabled in high-security contexts.
AI in Cybersecurity: A New Arms Race
This incident is not an isolated case. Over the past year, multiple research groups have demonstrated AI models capable of automatically finding vulnerabilities in widely used software. OpenAI and Anthropic have published papers on AI agents that plan and execute hacking sprees, autonomously chaining exploits across networks. The release of advanced models into the public domain—including through open-weight distributions—means that the capability to conduct such research is no longer confined to a handful of elite cybersecurity firms.
What A Security’s work on Zoom reveals is a shift in the nature of cyber risk. It is no longer sufficient to assume that obscure, low-level features in closed-source software are safe simply because they are hard to manually inspect. AI-driven fuzzing, static analysis, and reverse engineering have matured to the point where any function, no matter how esoteric, can be automatically probed with surgical precision. The democratization of these capabilities means that vulnerability discovery is becoming a commodity—and that the defense side must embrace the same tools at even greater speed.
The annotation protocol in Zoom was, in retrospect, a textbook target. It was proprietary, complex, and not subjected to the open review that cryptographic protocols or core networking stacks receive. AI models, trained on countless past vulnerabilities, instinctively looked there. The lesson for every software vendor is clear: if you have a feature that is rarely audited but deeply integrated into the user experience, it is likely carrying hidden bugs—and those bugs will not remain hidden for long.
What Users and Enterprises Should Do Now
For individual users, the immediate step is to update Zoom to the latest version. Zoom’s autoupdate mechanism should push the patch, but users can manually check for updates in the application menu. Enterprise administrators should enforce updates centrally through their device management platforms and consider whether annotation can be disabled for external meetings or low-trust scenarios. Even after patching, the broader lesson applies: video conferencing software is now a primary attack surface, and its features—screen sharing, chat, file transfer, annotation—should be audited with the same rigor as email or web browsers.
From a strategic perspective, organizations would do well to accelerate their adoption of AIa-assisted security tooling. The same models that discovered this Zoom bug can be deployed internally to scan custom applications, third-party integrations, and even legacy code that has not been updated. Waiting for a vendor to patch is no longer sufficient when the average time to exploitation after discovery can be measured in days or hours.
The Broader Implications for Software Trust
Zoom built its market dominance partly on ease of use and reliability. Users trust that joining a meeting is safe. That trust is now being stress-tested by a new class of threats that require no user error. This vulnerability did not rely on a weak password or a misclicked link; it relied on the very functionality that makes collaboration possible. As AI tools make such flaws easier to find, the trust landscape will shift. Software vendors will need to demonstarate not only that they fix bugs quickly, but that they proactively hunt for them using the most advanced tools available—and publish their methodology.
The A Security team’s choice to use publicly available AI models is telling. They did not require proprietary systems or massive compute clusters. The models they used are accessible to anyone with an internet connection. The message is unmistakable: the cyber security community must rethink its assumptions about which software is safe, how quickly vulnerabilities can be found, and who can find them.
A Race Between Defenders and Attackers
This Zoom disclosure is a preview of a future in which AI-driven bug hunting becomes routine. For defenders, the opportunity is to conduct continous, automated audits that cover every line of code and every protocol interaction. For attackers, the same capability means that zero-day discovery is no longer the exclusive province of nation-states or elite hacking groups. The cost of finding a critical vulnerability is falling toward zero.
What remains to be seen is how the software industry will respond. Will vendors start embededing AI-powered security testing into their development pipelines as a standard practice? Will users demand transparency about how often such audits are performed? Will regulators require it? The Zoom bug might be fixed, but the systemic vulnerability it highlights—the gap between the complexity of modern software and the human capacity to review it—will not be patched overnight.
In the meantime, users should update their Zoom clients today. The screen you share tomorrow may be the one you cannot afford to lose.
