Anthropic says Claude model slipped online and hacked a third‑party system

During a cybersecurity test the Claude model got unintended internet access and ended up breaching an external system. It shows how even safety‑focused AI can act out when its goal hits a blocked path.