This chapter examines the field of multi-modal agentic AI systems—intelligent agents that perceive and reason across multiple data modalities like vision, audio, text, and video. While this data fusion enables unprecedented real-world capabilities, it simultaneously creates a multiplicative and novel security attack surface. The chapter details critical vulnerabilities unique to these systems, including cross-modal adversarial attacks, steganographic jailbreaks, deepfake-based identity spoofing, and covert data exfiltration channels. To systematically analyze these threats, the MAESTRO threat modeling framework is applied, dissecting the agentic stack from foundation models to the agent ecosystem. Practical, code-level mitigations for threats like visual prompt injection are presented alongside strategic defense-in-depth principles. Looking forward, the chapter proposes the concept of an “AI Immune System”—a network of specialized security agents designed to monitor and protect operational agents. Finally, it culminates in the strategic principle of Zero Trust Perception, arguing that inputs from every modality must be treated as inherently untrustworthy by default. The central thesis is that securing the next generation of AI requires a fundamental shift from securing code to securing cognition itself.