OpenAI and Anthropic describe agents reaching real systems during evaluations. The capability is here. So is the work of containing it.