STORY-TRÅD·teknik

Top AI coding agents can be easy victims to sandbox escapes, showing they aren't as secure as they claim to be

What if a host component outside the sandbox reads AI coding agents' output? And what if that output is manipulated?