AI Security

An AI model operates inside a larger system. Training pipelines, retrieval sources, APIs, plugins, agents, user interfaces, identity controls, and cloud services can all affect its security. Protecting the model alone leaves other paths open to misuse or compromise.

The system boundary matters because the same model can present different risks in different deployments. A model that only summarizes public text has a different exposure profile from an agent that can query private data and execute tools.

AI security risks can arise from hostile inputs, untrusted training data, exposed model files, weak authorization, insecure tool use, and sensitive information in model outputs. Controls address these risks through input handling, access restrictions, isolation, monitoring, testing, and incident response.

No fixed control list applies to every AI system. A risk assessment should account for the model source, data sensitivity, deployment architecture, available tools, user population, and consequences of an incorrect or unauthorized action.

• Prompt injection and unsafe tool invocation
• Training data poisoning and model tampering
• Sensitive data exposure through inputs, retrieval, or outputs
• Unauthorized access to models, agents, or connected resources

AI security focuses on threats such as unauthorized access, manipulation, leakage, and abuse. AI safety addresses harmful behavior and failures even when no attacker is involved. AI governance sets the policies, roles, and oversight processes used to manage AI systems.

These disciplines overlap in areas such as testing, monitoring, and access control. Keeping the terms separate helps teams assign responsibility and select controls for a defined risk.