Comment by cubefox

Wrong dichotomy. During deployment, in actual production, users will also not have perfect sandboxes. So saying that the only problem was insufficient AI company security misses the point. Alignment is important, agents doing unintended things (even "going rouge") is a real problem.