OpenAI agents reached Hugging Face production during an evaluation
The companies described a containment failure involving a previously unknown vulnerability, reduced cyber refusals and thousands of agent actions.
Read moreSections
Research
Latest: Research. Coverage of AI models, products, research, chips, business and policy, with clear context and links to primary sources.Anthropic says safety tests crossed into three real systems
A review of 141,006 evaluation runs found six runs with external effects, including a malicious package that was briefly public.
Read more