A newly disclosed vulnerability in Zoom has highlighted a growing reality in cybersecurity: dangerous flaws can now be discovered with publicly available AI models in less time than it takes to schedule a patch window. The flaw, a zero-click remote code execution vulnerability in Zoom’s annotation feature, allowed attackers to take over the devices of participants on a call. Researchers say they found and exploited the bug with fewer than 20 AI prompts in under 24 hours.
The vulnerability was reported by a cybersecurity company called A Security. According to the firm, the problem existed in every version of Zoom across every operating system, including mobile, desktop, and web clients. Even calls protected with end-to-end encryption were not safe, which makes the bug especially concerning for high-profile uses such as board meetings, legal consultations, and healthcare appointments.
Zoom is one of the most widely used video conferencing platforms in the world, with millions of daily active users. The company has faced serious security problems before, including Zoombombing and other widely publicized incidents. But this vulnerability is especially threatening because it does not rely on a weak meeting password or a careless user. It is a technical flaw in the way the platform processes data, and it affects even the most security-conscious meetings.
The annotation feature became an attack surface
Zoom’s annotation feature is designed to let participants draw, highlight, and add text on top of a shared screen. It is a trusted, interactive tool used in classrooms, corporate meetings, and webinars. But under the hood, the client automatically parses whatever it receives when the annotation feature is active. This trust in incoming data became the vulnerability.
An attacker could send a specially crafted message to the receiving client, corrupting its memory and executing arbitrary code on the target’s machine. Because the protocol inside Zoom creates a direct channel between the person sharing the screen and each viewer, the attacker could target participants individually. This is not a phishing link or a malicious file. It is a zero-click exploit, meaning no action from the victim is required. Simply being on a call while annotation is enabled was enough to be compromised.
Zero-click vulnerabilities are the most dangerous type of bug in modern software. They bypass the common security advice to think before you click because there is no click. The victim does not have to open an attachment, visit a website, or approve a permission prompt. This makes detection and prevention significantly harder for both users and security teams.
The role of AI in finding the flaw
What makes this particular disclosure notable is the way the vulnerability was discovered. The researchers at A Security said they used publicly available AI models to find the flaw. They did not rely on a frontier, restricted-access model from one of the top AI labs. The process took fewer than 20 prompts and less than 24 hours, according to the firm.
That is a startlingly fast timeline. For years, vulnerability researchers have used automated tools to fuzz protocols, scan source code, and identify suspicious patterns. But AI-driven vulnerability discovery changes the game in a fundamental way. Instead of laboriously tracing code paths and reverse-engineering proprietary protocols, a researcher can ask a model to identify potential memory corruption spots, generate test cases, and even suggest an exploit strategy. The model still needs human guidance, as A Security emphasized, but the barrier to entry has dropped dramatically.
The speed of this discovery has renewed concerns about the cybersecurity landscape. Just last week, officials from the United States and the United Kingdom warned at the Black Hat cybersecurity conference in Las Vegas that the pace at which vulnerabilities are being discovered is quickly surpassing the ability to patch them. When a bug can be found by an AI model in a single workday, but enterprise IT teams need weeks to test and deploy patches across thousands of endpoints, the window of exposure is no longer measured in months. It is measured in hours.
Why public AI models are enough
It is common for the most advanced AI systems to be gated behind waitlists, usage caps, and safety filters. The frontier labs have argued that access to their most powerful models must be restricted so that the tools do not fall into the wrong hands. But this disclosure suggests that even the models available to anyone with an internet connection are powerful enough to cause major disruption.
The implication is not that frontier models are overhyped, but rather that the state of the art has already reached a threshold of practical usability for malicious tasks like vulnerability discovery. The AI does not need to understand the entire codebase of a complex application. It simply needs to assist a competent researcher in narrowing down likely weak points and generating exploit code.
This aligns with warnings from cybersecurity experts over the past year that AI would lower the skill ceiling for hacking. You no longer need to be a memory-corruption expert to find a memory-corruption bug. You need to know the right questions to ask an AI and how to evaluate its suggestions. That is still a meaningful skill, but it is far more common than deep low-level systems expertise.
The challenge for patching and response
Zoom has since patched the vulnerability, urged users to update their clients, and said no evidence of active exploitation was found. But the patching process itself illustrates the problem. Even when a vendor ships a fix, every user must update. That sounds simple, but many enterprise environments still run outdated Zoom versions because of compatibility requirements or because IT teams have not enforced automatic updates.
The vulnerability’s presence in end-to-end encrypted calls also complicates the story. E2EE is generally promoted as the gold standard for security, as it prevents Zoom and third parties from reading call contents. But as this bug shows, encryption only protects data in transit. It does not protect the client software running on your device. The exploit targeted the parsing of annotation data, not the media stream itself. Even encrypted connections can hide malicious payloads from network inspectors because the payload travels encrypted. In this case, the encryption likely made the exploit harder to detect, not easier.
For individual users, the immediate action is to update Zoom to the latest version. On most platforms, Zoom can be configured to update automatically, but these settings are not always enabled by default. Enterprise administrators should check their deployment policies, push the update through their management console, and verify that legacy clients are no longer allowed to connect. In critical environments, it may also be wise to disable the annotation feature completely until all endpoints are known to be patched.
The disclosure also raises questions about the broader ecosystem of video conferencing and collaboration tools. If Zoom’s annotation protocol could be broken by a small team with a few AI prompts, it is fair to assume other products have similar weaknesses. Many collaboration tools feature screen sharing, co-browsing, and real-time interactive annotations, all of which require clients to parse data from unknown or semi-trusted sources. The attack surface is enormous.
The new normal for cybersecurity
The discovery of this vulnerability at the hands of AI models is not an isolated event. Over the past year, dozens of security firms and academic researchers have demonstrated AI agents that can autonomously audit source code, identify dependencies with known vulnerabilities, and craft exploit scripts. Some of these efforts have produced legitimate results, while others have been used in actual attacks against schools, gyms, and government agencies.
The trend likely points to an arms race. Defenders will use AI to scan more code faster and automate patch deployment. Attackers will use AI to find flaws that are invisible to traditional tools. The side with the better AI, faster feedback loops, and smarter human oversight will have an advantage. But the scale and speed of AI-driven discovery are already stretching the limits of current security practices.
The disclosure is a reminder that safety was never guaranteed. Modern software is complex, patching is slow, and the tools available to researchers and attackers are becoming more powerful every month. We are all increasingly vulnerable without fully realizing it, and the pace of vulnerability discovery is only going to accelerate from here.
Source:Gizmodo News