Security
Most OpenClaw incidents do not look dramatic at first. They look like defaults, convenience shortcuts, and one more thing you meant to clean up later. These are the mistakes I keep seeing.
Mistake 1
If your gateway is bound to 0.0.0.0 without a very clear reason, you are increasing your exposure surface immediately. Loopback by default is the sane choice. Use a controlled remote path like Tailscale if you actually need outside access.
Mistake 2
Inline API keys in openclaw.json are not convenient. They are future incident reports. Use a secrets provider, lock down file permissions, and assume logs and screenshots will eventually betray you if you get lazy.
Mistake 3
A conversational agent should not have the same blast radius as a coding agent or an ops agent. If every role can touch everything, your mistake budget is tiny.
Mistake 4
A secure system is not just hard to break into. It is also hard to break accidentally and easier to recover when it goes sideways. If compaction, restart, or sleep-wake events leave you guessing, the setup is not production-safe.
Mistake 5
If you cannot reconstruct what happened, what the agent was allowed to do, and what actually ran, you are operating on vibes. That gets expensive fast.
Send your config and I will give you the highest-risk findings, the fastest fixes, and the parts that can wait.
Request the free audit