AI is changing all aspects of life, from healthcare to finance and transportation. The more you rely on it, it’s exposed to considerable cybersecurity threats. This research investigates significant vulnerabilities in AI systems, specifically, adversarial attacks, Large Language Model (LLM) security, and malicious data in training. Manipulation can be made by adversarial techniques such as evasion, poisoning, and backdoor attacks on AI models, causing it to deliver inaccurate and harmful outputs. Besides, LLMs have issues like prompt injection, data leakage and insecure APIs, which exploit sensitive information. Moreover, cybercriminals find ways to exploit the weaknesses in AI models through data poisoning, label flipping, and injection attacks when it comes to training data security. To tackle these challenges, this study looks at defensive strategies, including adversarial training, anomaly detection, secure input/output pipelines, data validation, etc. Like robust security measures, reusing a system across its lifecycle can increase reliability and trustworthiness, making it more robust and resilient to security threats. The objective of this research is to examine the risks associated with AI implementations and suggest mitigation plans to be incorporated into different stages of the AI lifecycle. The research proposes a comprehensive framework with enhanced reliability and robustness to protect AI systems against ever-evolving cybersecurity threats.