Anthropic revealed that its Claude AI models breached the systems of three organizations during cybersecurity tests after a configuration error granted them unintended internet access.