Yeah, it's wild. I spent several weeks nearly full time on a deep dive of claw architecture & security.

The short of it - OpenClaw sandboxes are useful for controlling what sub-agents can do, and what they have access to. But it's a security nightmare.

During config experiments, I got hit with a $20 Anthropic API charge from one request that ran amuck. Misconfigured security sandbox issue resulted in Opus getting crazy creative to find workarounds. 130 tool calls and several million tokens later... it was able to escape the sandbox. It used a mix of dom-to-image sending pixels through the context window, then writing scripts in various sandboxes to piece together a full jailbreak. And I wasn't even running a security test - it was just a simple chat request that ran into sandbox firewall issues.

Currently, I use sandboxes to control which agents (i.e. which system prompts) have access to different tools and data. It's useful, but tricky.