AI Trust Glossary · Canonical Definition
Sandboxing
Running an AI agent in an isolated execution environment with restricted permissions - limiting what actions it can take and what data it can access.
Explanation
Sandboxing enforces least privilege at the execution level. An agent in a sandbox can only access systems and data it has been explicitly granted. Sandbox escape - gaining access to resources it should not have - is a critical security failure.
Why it matters
Unconstrained AI agents with broad system access represent catastrophic risk if they malfunction or are compromised. Sandboxing limits blast radius. An agent that hallucinates or is prompt-injected in a proper sandbox causes limited damage; one with unrestricted access can cause unlimited damage.
How Borealis uses it
Sandboxing is a recommended deployment practice that directly supports constraint adherence. Agents deployed with proper sandboxing have a structural enforcement layer reinforcing behavioral constraints. Audit evidence should include sandbox configuration to demonstrate infrastructure-level constraint enforcement.
See also