Sandboxing
Also called: isolation, containment
Running an agent, tool, or MCP server in a constrained environment so a compromise cannot reach the host or other components.
Sandboxing ranges from OS-level (containers, VMs, seccomp, restricted users) to network-level (egress filtering, no access to the metadata endpoint or internal services) to MCP-level (isolating each connected server's context and credentials from the others). The goal is that "the agent got prompt-injected" does not become "the host is compromised" or "every other server's secrets are gone."
Why it matters
Because you cannot reliably stop the injection, containment is where a lot of real safety comes from. A sandboxed local MCP server that gets rug-pulled can still only damage what is inside its box. Sandboxing is most effective combined with least privilege — a small box with few keys in it.
Exposure map
Live counts of instances showing this pattern will appear here once the exposure map is collecting data.