Zoom has quietly patched a critical security vulnerability that could have allowed an attacker to take over every device connected to a meeting. The flaw, which researchers have nicknamed “Zoomsday,” was discovered using an unconventional method: fewer than 20 prompts on publicly available AI models. The disclosure has sent shockwaves through the cybersecurity community, as it demonstrates how quickly artificial intelligence has lowered the barrier to discovering and exploiting serious software bugs.
The vulnerability lived inside Zoom’s annotation feature, a widely used tool that lets meeting participants draw or write on shared screens. The feature is popular in classrooms, boardrooms, and remote support sessions, making it an attractive target for malicious actors. According to the security firm that discovered the flaw, an attacker could join or host a meeting and trigger the exploit without any action from victims. Once executed, the malicious code could steal passwords and other sensitive data, activate the camera or microphone without consent, or install ransomware and other persistent malware. Critically, the attack left no visual cue — no glitch, no notification, no suspicious pop-up — so even the most careful users would have been completely unaware that their device had been compromised.
Security researchers believe this is one of the first major real-world cases where a working exploit was generated largely by an AI agent. Traditionally, finding a bug like this and turning it into a weaponized exploit required deep expertise in reverse engineering, memory corruption, and platform-specific internals. Highly skilled teams would spend weeks or months analyzing code, fuzzing inputs, and crafting payloads. Governments and defense contractors treated such capabilities as strategic assets. But now, researchers say, that kind of work can be accomplished in a single day using off-the-shelf AI models and carefully chosen prompts.
How the Exploit Works
Zoom’s annotation feature renders shared content in real time, with each participant receiving drawing updates from others. This involves parsing complex data structures that contain coordinates, colors, transparency settings, and other visual attributes. A vulnerability in this parsing logic could allow an attacker to supply specially crafted input that triggers a memory corruption bug, leading to arbitrary code execution.
The exact technical details of the exploit have not been fully disclosed, but the security researchers described a chain of weaknesses that could be chained together. The first step was identifying a memory-unsafe operation in the annotation rendering path. From there, the attacker could overwrite critical pointers or use a technique known as “use-after-free” to gain control of the program’s execution flow. Finally, a payload could be delivered that runs with the privileges of the Zoom application, which under normal conditions has broad access to the camera, microphone, file system, and network.
When the researchers prompted the AI model, they asked it to analyze publicly available code snippets, look for patterns associated with past Zoom vulnerabilities, and suggest potential attack surfaces. The model generated hypotheses that the team verified and refined. Within a few hours, the AI produced a step-by-step exploitation plan that human researchers then converted into a working proof-of-concept. The fact that this process required fewer than 20 prompts means that someone with little more than a web browser and a subscription to a commercial AI service could potentially replicate the discovery.
AI-Powered Hacking Is Here
The “Zoomsday” discovery illustrates a dramatic shift in offensive cybersecurity. For years, AI was seen as a useful tool for writing phishing emails or automating simple network scans. But large language models have become far more capable at understanding code logic, spotting unusual patterns, and even generating low-level assembly instructions. Security researchers have shown that modern models can solve capture-the-flag challenges, reverse-engineer binaries, and identify vulnerabilities in open source repositories. Applying that capability to a closed, widely deployed application like Zoom was a significant next step.
Idan Levcovich, one of the lead researchers on the project, noted that producing this exploit would have been considered “nation-state work” just a few years ago. Elite teams with substantial budgets and government backing were funded to develop exactly this kind of capability. The fear now is that the same technology is available to independent hackers, criminal enterprises, and even script kiddies. If a single day is enough to find a serious vulnerability in a major platform, the cost of discovering new zero-days could drop to almost nothing.
This has major implications for the way software companies approach security. Traditional bug-bounty programs reward researchers who report flaws, but if AI can find bugs faster, companies will need to invest more in automated testing, memory-safe programming languages, and constant security audits. It also raises the risk that malicious actors could use AI to discover vulnerabilities before vendors have a chance to patch them. A “patch gap” of even a few hours could be enough for a large-scale attack.
Zoom’s Security History
Zoom has faced several high-profile security and privacy controversies over the years. In 2020, the platform became a household name as the pandemic forced millions to work, learn, and socialize remotely. That rapid growth brought scrutiny. Researchers discovered “Zoombombing,” where uninvited users disrupted meetings by sharing offensive content. There were also flaws that allowed attackers to join meetings without authorization or steal Windows login credentials.
One of the earliest serious incidents involved a vulnerability in Zoom’s Windows client that could allow an attacker to steal the user’s NTLM password hashes simply by opening a malicious link. Another issue allowed websites to enable a user’s camera without permission if the user had installed an older version of the Zoom app. Zoom responded by launching a 90-day security initiative, hiring prominent security experts, and adding end-to-end encryption for calls. However, the company continued to face vulnerabilities in its codebase.
This latest flaw is particularly concerning because it affects the annotation feature, which is used by teachers, product teams, and customer support agents. The fact that the exploit requires no user interaction means that a malicious meeting participant could silently compromise every other attendee. It also works across all major operating systems, making it one of the most severe Zoom flaws to date.
Lessons for the Industry
The discovery of “Zoomsday” is a wake-up call for the entire software industry. It shows that the threat landscape is entering a new phase where AI is not just an assistant but a driver of vulnerability research. Companies need to rethink their threat models. A single researcher with an AI tool can now perform what used to be a team of highly paid experts’ work. This democratization of exploit development is both empowering and dangerous.
For defenders, the same AI tools can be used to harden applications. Automated code review, vulnerability scanning, and even automated patching are becoming more practical as language models improve. But the attackers also have access to these tools. The race between offense and defense is likely to accelerate, and the winners will be determined by who can integrate AI into their workflows most effectively.
The “Zoomsday” disclosure also highlights the importance of coordinated vulnerability disclosure. The security firm did not publish the exploit details until Zoom had issued a patch. That is the responsible approach. Once the patch is applied, the vulnerability is no longer exploitable, even if the details become public. The researchers did not provide the specific AI prompts they used, but they demonstrated the potential for AI-assisted hacking in a way that will shape security research for years to come.
Zoom users are advised to update the application immediately. The company has rolled out patches for Windows, macOS, Linux, Android, and iOS. The update process is usually automatic, but users on managed devices may need to contact their IT department to ensure the patch is applied. Since the attack leaves no trace, it is especially important to verify that the Zoom version in use includes the security fix.
The security researchers behind this discovery are not simply sharing a story of a clever hack. They are announcing that the era of AI-powered vulnerability discovery has truly begun. What was once the domain of elite nation-state operations is now accessible to anyone who can type a prompt. This should prompt every organization that relies on Zoom — or any similar platform — to review its security posture and assume that AI could be used against them. The future of software security will depend on how well both companies and researchers harness the same powerful technology that exposed this flaw.
Source: The Verge News