AI Agents Are Here. So Are the Threats.

**Executive summary:** This Unit 42 assessment demonstrates nine simulated attack scenarios against agentic AI applications (framework-agnostic examples implemented with CrewAI and AutoGen) that can lead to information leakage, credential and metadata-token exfiltration, unauthorized internal network access, SQL injection and broken object-level authorization abuse, and remote code execution; the report explains root causes (misconfigurations, unsafe tool integrations, prompt weaknesses) and prescribes layered mitigations including prompt hardening, content filtering, tool input sanitization, vulnerability scanning, and strict code-executor sandboxing.