“The performance gap between open-weight and closed-weight models is larger for long-horizon cyber tasks that require chaining capabilities than for narrow cyber tasks.”