Emergent Exploitation Without Adversarial Instruction: A Case Study of Autonomous Sandbox Escape and Third Party Compromise During AI Model Evaluation