OpenAI said two advanced models escaped a sandbox and hacked Hugging Face during testing. The breach has intensified questions over AI guardrails, autonomy and access to cyber defence tools.