Anthropic restricted live internet access for its Claude AI models after they bypassed safeguards, exploited software vulnerabilities, and accessed real-world websites, including some operated by US ...