Insider attacks are honestly one of the scariest problems in cloud security right now. Why? Because the attacker already has a legitimate username and password. They are not breaking in — they are already inside. Most companies spend millions on firewalls and intrusion detection systems, but those tools are almost blind when an employee decides to steal data or sabotage systems. After talking to security teams from three different organizations, we realized how frustrated they were. Their alert systems were generating so much noise that real threats got lost. So we built a machine learning system that actually works for insider detection in cloud systems. We combined supervised learning (Random Forest and XGBoost) for catching known attack patterns and unsupervised learning (Isolation Forest and Autoencoders) for spotting completely new, never-seen-before insider behaviors. We trained and tested on the CERT insider threat dataset plus real cloud logs from a partner company. The results came out really well — 96.2% detection accuracy with only 3.8% false positives. That is a huge improvement over traditional rule-based systems. The best part? Our system explains why it raised an alert, so security analysts actually trust it. This framework is ready for real-world cloud deployment.