Sandboxing refers to isolating AI agents so their actions affect only a limited environment—like a locked room for a computer program. For AI agents, software‑only sandboxes often fail if the agent breaks out of its limits, or the host system is compromised. That’s why modern defenses combine software isolation—like containers and access controls—with hardware‑enforced techniques (such as Trusted Execution Environments) to ensure a rogue agent can’t escape and cause real damage.