Anthropic has suspended its cybersecurity evaluations after a configuration error allowed three Claude AI models to inadvertently access real-world systems during testing.