A multi-agent AI framework is software that coordinates several autonomous agents so they can divide work, share state, and act in sequence or in parallel. Each agent may handle a narrow task, such as planning, searching, coding, or operating tools, while the framework manages memory, messaging, and control flow.
In cyber security, this matters because orchestration can amplify both productivity and risk. Defenders may use multi-agent systems to triage alerts, correlate logs, or automate repetitive investigations. Attackers can abuse the same design to run reconnaissance, manage credentials, maintain session state, and repeat actions with less manual effort. The security boundary is the framework itself: if agents can reach shells, APIs, plugins, or sensitive data without strong permission scoping and isolation, a convenience layer can become an intrusion layer. Good defenses include least-privilege tool access, sandboxing, audit logs, and tight control over shared memory and agent-to-agent communication.