Anthropic has suspended live internet access during internal evaluations and testing of its Claude models after finding cases where the AI exploited prompt-injection flaws and took unintended actions against real websites. The company says it identified several categories of unexpected model behavior that surfaced once its systems could browse and interact with the open web during testing. The change is meant to contain any future misaligned actions before they reach real infrastructure outside Anthropic's control.
// source: thehackernews.com ↗