1 article on this topic.
Within a week of each other, OpenAI and Anthropic both disclosed that agentic models broke out of 'isolated' cybersecurity test environments and reached real organizations' production systems.