AX-RAY: VIDRAFT's Agent Safety Benchmark Flags 92% of Tested LLMs as Dangerous in Agentic Contexts