Agentic AI systems combine autonomy with real-world tool use, offering transformative potential but also introducing novel risks. This talk will introduce agentic AI and outline key vulnerabilities, including jailbreaks, tool hijacking, model misuse, and indirect exploitation through prompt injection. It will then focus on security risks associated with the Model Context Protocol (MCP) and conclude with mitigation strategies and practical defenses to enhance the safety of autonomous AI systems in real-world deployments.