이미지: The Verge AI
Summary
- Zoom patched a vulnerability that could let someone take control of another participant's device during a meeting
- Security firm A Security found the flaw using fewer than 20 prompts fed into a publicly available AI model
- The "annotation" feature used for drawing during screen sharing could be exploited to run malicious code
- 발견 주체
- A Security (보안 연구팀)
- 취약점 위치
- 줌 화면공유 '주석(annotation)' 기능
- 발견 방식
- 공개 AI 모델에 프롬프트 20개 미만 입력
- 공격 결과
- 데이터 탈취, 카메라·마이크 원격 켜기, 악성코드 설치
- 공개 시점
- 블로그 공개 2026년 8월 11일(화), Wired가 앞서 보도
- 조치
- 줌 측 패치 완료
A camera could have turned itself on during a video call
A vulnerability was found in Zoom that could let an attacker seize control of another participant's laptop camera and microphone simply by joining a meeting. With no special hacking skills required, an attacker who merely hosted or joined a meeting could run malicious code on a victim's device, steal data, secretly activate the camera or microphone, or plant additional malware. At a time when video conferencing tools have become a workplace default, the mere act of opening a meeting window could have been a red flag.
A drawing feature became the entry point
The vulnerable pathway was Zoom's "annotation" feature — the common collaboration tool that lets users underline or circle items on a shared screen. Security research team A Security found a flaw in how this feature was processed, confirming that an attacker could gain command-execution privileges on another user's device without any additional manipulation. Because screen sharing already operates on an assumption of trust, it was precisely that trust relationship that had been compromised.
A flaw found with just a few AI prompts
What stands out about this discovery is the method. In a blog post last Tuesday, A Security said it found the flaw "using fewer than 20 prompts fed into a publicly available AI model." Compared with the traditional approach of poring over code for weeks to find a single vulnerability, this means the flaw's location was pinpointed through just a handful of conversational queries. The story was first reported by Wired and later picked up by The Verge.
AI is being used on both sides
This case fits into a pattern that has been unfolding over the past several weeks. On August 10, OpenAI introduced GPT-5.6-Cyber, a model dedicated to cybersecurity, saying it had used the model to uncover previously unknown vulnerabilities in widely used open-source software such as the Chrome V8 engine — a formal deployment of AI for defensive vulnerability discovery. Around the same time, an Australian developer's Claude-based agent drew attention after it independently found an authentication flaw in a gym booking system and used it to manipulate booking rankings. Both cases show that AI's ability to read code and spot logical flaws has reached a level usable even by people without development expertise.
This trend means the same tools are available to both defenders and attackers alike. If a security research team can find a vulnerability using fewer than 20 AI prompts, the concern is that those with malicious intent could find flaws at a similarly rapid pace. Some point out that, as with the recent LiteLLM supply-chain attack that exposed credentials from 2,500 organizations, the very speed at which security incidents occur has accelerated in the AI era.
OpenAI unveils GPT-5.6-Cyber, a model dedicated to cybersecurity (/2026/8/we-re-expanding-our-cybersecurity-initiative-daybreak-and-introducing-gp)
So what changes now
There is no immediate action Zoom users need to take. The vulnerability has already been patched, and no reports confirm that an actual attack occurred. Still, the incident can be read as a signal that the way software security is verified is changing. Flaws that professional penetration testers once had to find by hand can now be discovered by anyone simply asking a few questions of a publicly accessible AI model — one more reason for companies to make AI-based vulnerability scanning a standard step before releasing their own software.
