Generative AI and AIOps are increasingly embedded in enterprise observability, automated remediation, service management and the software supply chain, which widens the attack surface from prompts to training data and model updates. This work proposes and evaluates a multi layered defense against prompt injection, data poisoning, adversarial input manipulation and supply chain tampering. A controlled dataset of 172 attack instances was constructed, with an equal number of instances for each threat class. Every instance was run against an unsecured baseline, an input gateway configuration and a full defense configuration that combines data screening, provenance checks, adversarial validation, policy enforcement, continuous monitoring and incident response. The experimental workflow was implemented in Python using a structured security test harness, with Matplotlib for visual analysis and logging utilities for reproducible evidence capture. The results show that although isolated controls can stop attacks that are clearly malicious, vulnerabilities remain at cross layer points and are exposed when attacks exploit trusted context, polluted training material or operational automation. The layered approach further reduces the likelihood of successful attacks and preserves useful model behavior while improving the ability to detect, contain and recover from an attack. The paper therefore recommends a practical security pattern for organizations deploying generative AI and AIOps components in production environments, where governance, auditability and controlled automation must work together rather than as standalone security considerations.

Prompt injection data poisoning adversarial attacks generative AI AIOps