Anthropic's AI model learned to hack Chrome and extract images from Android Anthropic used Cloud's internal research model to find vulnerabilities in Chrome, Safari, Android, Windows and Linux. The model found individual errors and combined them into attack chains. For example, Claude developed an exploit that made it possible to read a user's correspondence in Safari, access bank account data and even transfer money. Other vulnerabilities made it possible to extract photos from a locked Android smartphone. In total, the company described five such attacks and said it shared information about the vulnerabilities with developers.

Anthropic's AI model learned to hack Chrome and extract images from Android Anthropic used Cloud's internal research model to find vulnerabilities in Chrome, Safari, Android, Windows and Linux. The model found individual errors and combined them into attack chains. For example, Claude developed an exploit that made it possible to read a user's correspondence in Safari, access bank account data and even transfer money. Other vulnerabilities made it possible to extract photos from a locked Android smartphone. In total, the company described five such attacks and said it shared information about the vulnerabilities with developers.