Long-horizon AI is an AI system built to operate across many steps, not just answer a single prompt. It can keep context, plan ahead, revisit earlier actions, and work toward a goal over time. This makes it more powerful than a one-shot chatbot, but also harder to fully predict or audit.
In cyber security, long-horizon AI matters because persistence changes the risk surface. A system that can make repeated decisions may interact with files, tools, APIs, or code over many stages, so defenders need tight sandboxing, least-privilege access, action logging, and human review gates. Attackers may abuse the same capability for multi-step phishing, reconnaissance, malware adaptation, or automated exploitation. Defenses also use long-horizon AI for tasks like incident triage, threat hunting, and proof checking, but those outputs still require verification before any sensitive action is taken.