OpenAI, Anthropic AI Models Breached Systems During UK Safety Tests