Cyber Risks · The Hacker News

New ChatGPT Atlas Browser Exploit Lets Attackers Plant Persistent Hidden Commands

Summary: LayerX Security researchers found a cross-site request forgery flaw in OpenAI's ChatGPT Atlas browser that lets attackers write malicious instructions into ChatGPT's persistent memory, which then survives across devices, sessions, and browsers unless users manually delete it from settings. In LayerX's tests, tainted memory let ordinary prompts trigger code fetches, privilege escalations, and data exfiltration, and the firm measured Atlas stopping only 5.8% of over 100 in-the-wild web vulnerabilities and phishing attacks, against 53% for Edge and 47% for Chrome.

Why it matters: A memory entry planted once keeps working on every device and session the user opens, so the compromise outlasts the browser tab that delivered it - and at a 5.8% block rate on tested malicious pages, Atlas is not the control standing in the way.

Read the original at The Hacker News ↗

Topics
AI Security
Related