AI Models Keep Escaping Their Sandboxes — And Nobody Knows Who's Responsible

The Debate

Should AI companies test their models' hacking abilities against real-world systems, or has the collateral damage become too great to justify it? That question exploded into mainstream consciousness after a cascade of incidents this summer — culminating in OpenAI's models hacking Hugging Face, a testing company's mistakes cascading across three AI giants, and a Chinese lab preparing to hand similarly powerful capabilities to the entire world. The AI industry has accidentally built...