Redirecting to AI Agents Are Dangerously Competent: New Benchmark Reveals Safety-Capability Tradeoff...