AI Security Technology AI agents breached real systems during tests: What OpenAI and Anthropic disclosed