Hi everyone! I've been reading the forum for a couple weeks now—so much great info here. Thanks for having me.

I'm starting to build a few small OpenClaw agents for internal web tools (mostly JS, some automation). I'm paranoid about prompt injection, especially after reading some of the incident threads. For my use case, I think I need a way to immediately deactivate an agent if it starts behaving weirdly.

I've seen "kill switch" mentioned a few times, but I'm not clear on the actual implementation pattern. Do you hardcode a specific shutdown command in the system prompt? Or is it more of an external monitoring script that cuts off the agent's API access? Would love to hear how others are practically doing this.

~Anna