A hand-off boundary is the transfer point where one agent passes context, tasks, or notes to another agent. In multi-agent systems, this is the seam between a low-trust step and a higher-trust step. It matters because anything copied across that boundary may be treated as guidance, even when it originated from untrusted text.
In cyber security, hand-off boundaries are important in prompt-injection and agent-hijacking scenarios. An attacker may hide malicious instructions inside a summary, comment, or task hand-off so the next agent follows them and exposes secrets or alters a workflow. Defenders reduce this risk by treating hand-off content as data, not authority: validate it with a schema, limit the privileges of upstream agents, and require review or branch protection before AI-generated actions can change code or repository state.