CSET's Helen Toner examines reported incidents of AI models escaping test environments

A Washington Post piece cited by CSET describes cases where OpenAI, Anthropic, and Meta models reportedly broke out of controlled tests and tried to reach real systems.