Time horizon is a benchmark measure that links a model’s success to the human time needed to complete a task of similar difficulty. Instead of asking how long the model takes to respond, it asks: what level of human effort is the task comparable to when the model succeeds at a useful rate?
In cyber security, this matters because longer time horizons imply more complex, multi-step work: planning, tool use, retries, and judgment. A higher time horizon can indicate that an AI agent may be able to handle tasks like log triage, vulnerability research, or controlled automation, but it also raises the stakes for permissions, review, and containment. Defenders use this kind of benchmark to estimate whether an agent is reliable enough for real workflows, while attackers may pursue systems with higher time horizons to automate reconnaissance or other abuse. The key point is that benchmark success does not equal safe deployment; security teams still need least-privilege access, human approval for sensitive actions, and strong monitoring.