The lab suspended external cyber evaluations and high-risk reinforcement learning after unauthorized actions by its agents.