I took the brand new Mac Mini (M4, 16GB RAM) and tried to push it into something I honestly don’t trust yet: a fully local AI assistant that can read, reason, and act. I ran OpenClaw (formerly “Clawbot/Cloudbot”) on top of Ollama with a local Qwen 2.5 model, because my whole goal was to keep things off the cloud and away from my accounts. I even made a separate non-admin macOS profile ("Jarvis") with zero iCloud access, because giving an agent control over a machine tied to your real life is… yeah, no thanks. But the more I set it up, the more it turned into a security deep dive. Running local helps privacy, but it doesn’t automatically make the system secure—prompt injection and “trust confusion” are still a thing. I also ran into the reality of agent overhead: my model could answer in ~1.6 seconds in terminal, but the same prompt inside OpenClaw took ~24 seconds because of all the extra layers happening in the background. Then I tested how dangerous this gets when an agent can touch your filesystem. My rules helped block some actions, but with direct commands the protection fell apart, and the agent even rewrote my Python script and wiped its contents. My takeaway: these tools are powerful, but they’re not hardened by default—if you don’t understand the permissions and controls you’re setting up, don’t run an agent on a machine you care about.