Cyber Risks · The Hacker News

ChatGPT Atlas Browser Can Be Tricked by Fake URLs into Executing Hidden Commands

Summary: NeuralTrust reported Friday that OpenAI's newly released ChatGPT Atlas browser can be jailbroken through its omnibox, which interprets input as either a URL or a natural-language command to the agent, letting a crafted prompt disguised as a URL-like string starting with "https" and a domain-like text be treated as high-trust user intent. Because the malformed input fails URL validation, the agent executes the embedded instruction and redirects the user to the site named in the prompt, which could sit behind a "Copy link" button leading to phishing pages or carry a hidden command to delete files from connected apps like Google Drive.

Bottom line: OpenAI's own CISO, Dane Stuckey, calls prompt injection a "frontier, unsolved security problem" even after extensive red-teaming and added guardrails, and the same vector has already hit Perplexity Comet and Opera Neon - so the browser agent your users adopt next is a trust boundary you do not control.

Read the original at The Hacker News ↗

Topics
AI Security
Related