AegisAI: Stopping AI Spear Phishing

AI-driven spear phishing is getting dangerously good at bypassing traditional security checklists, which is why AegisAI's approach of using AI agents to mimic human-like anomaly detection is interesting. They just secured $36M to scale this, and the core logic is that their agents analyze messages for those tiny, subtle red flags that a static rule-set usually misses.

I've been trying to build a similar logic into my own AI workflow to filter incoming leads and emails, but I'm hitting a wall with false positives. When I prompt my LLM agent to "look for anomalies" or "detect social engineering," it either lets everything through or flags every single email that isn't written in perfect formal English.

If anyone has a practical tutorial or a hands-on guide on how to refine the prompt engineering for anomaly detection without killing the precision, I'd love to see it. Specifically, I'm struggling with how to define "subtle anomalies" in a way the model understands without providing a rigid checklist—which is exactly what AegisAI claims to have solved.