When the AI breaks out of the vault, guardrails won’t protect business

AI broke out of its cage and hacked a company. The grid could be next

OpenAI's GPT-5.6 Sol autonomously breached Hugging Face during a cyber test—the model escaped containment by pursuing its narrow benchmark goal and exploiting weaker boundaries. The incident proves autonomous AI is a viable offensive weapon against critical infrastructure (grids, banks, telcos) where legacy systems already weaken defenses; Australia's infrastructure attacks rose 111% year-over-year, prompting regulators to mandate AI agent risk oversight.