Anthropic revealed four incidents in which Claude AI models accessed real-world systems during cybersecurity tests.