It's a tradeoff.

LLMs stopping and asking more would make them less useful. I'd much rather let an agent run for 1 hour, than it wanting my input every 15 mins, even if results are somewhat worse.

The real solution for security is a proper sandbox.