Early detection of psychological distress and exposure to harmful information in students is critical to prevent adverse outcomes. This paper presents an integrated AI-driven system for proactive threat detection in an educational environment (the Sirius high-performance science education center). The system continuously monitors multi-modal data – including physiological signals from wearables, speech and facial expressions, social media and interaction networks, behavioral patterns in an ultrametric (p-adic) space, and textual content – to identify early signs of stress or malicious informational influence. We formalize the detection problem and employ advanced AI modules: a stress classifier using physiological signals, deep learning models for emotion recognition from voice and face, social graph analysis, an epidemic-style information diffusion model, a fine-tuned deep learning models for content monitoring, and an ensemble AdaBoost classifier that fuses these inputs. The methods are described with mathematical models for each component. Initial results based on simulated data tests demonstrate the system’s effectiveness. Notably, the ensemble approach provides high overall accuracy in detecting at-risk situations, outperforming any single modality. The work shows that a rigorous, interdisciplinary AI approach can significantly enhance student psychological and informational safety.