The four gates framing is the clearest version of this I've read. "A page can't do anything by itself, it can only change what the agent decides next" is the sentence I'll use the next time someone asks if they just need to avoid shady sites.
The row that surprised me most was reads. It's easy to turn on a sandbox and assume your keys are safe, when it's really protecting writes and outbound traffic. Calling out that ~/.ssh is readable by default in every column is worth the whole table.
One habit that pairs well with your AGENTS.md block: split research and changes into separate sessions. Let the agent browse and write up what it found, then start a fresh session from that summary before it edits files or runs commands. The untrusted rule helps, but it still sits in the same context as whatever the page said. A fresh session means the injected text isn't there at all when the risky tool calls happen.
Have you found a clean way to fence in MCP servers that genuinely need network, like a GitHub one, without opening the whole network to them?