What’s your worst "AI Agent gone rogue" moment in production?

As we move rapidly from experimental chat interfaces to fully autonomous agents with direct API access, runtime security is becoming a massive blind spot.

While building Aegisora (our open-source, zero-latency runtime proxy), I constantly hear terrifying stories from engineering teams about agents trying to execute unauthorized API calls, looping endlessly, or silently leaking PII because static guardrails and prompts failed.

I want to hear your engineering war stories: • What broke when you gave your AI agent too much autonomy? • Did it try to drop a database, hallucinate a critical tool call, or rack up a massive API bill? • How did you catch it and fix the vulnerability?

Let’s share the chaos and talk about real production bottlenecks! 👇