Hey dev.to community 👋 Over the past few days, I’ve been running dozens of AI Agent experiments. The more powerful they got, the more nervous I became. They can do amazing things, but they’re also shockingly easy to jailbreak, abuse tools, or quietly exfiltrate data. And once they go rogue? Traditional “safety inside the agent” approaches just don’t cut it. So I decided to solve it differently. The Solution: Sentinel v0.3.0 “The Shield Release” I pulled the entire security layer completely outside the agent using an independent Shield Sidecar process. The agent literally cannot see it or kill it. Every risky action (shell commands, file I/O, API calls, etc.) must request permission from the Shield first.…