The AI attack surface is the full set of points where an artificial intelligence system can be abused, manipulated, or targeted. That includes training data, model inputs, prompts, APIs, plugins, output channels, and the infrastructure that hosts the model. In practice, it matters because AI systems often make decisions or generate text based on data they do not fully control, which can let attackers steer results, extract sensitive information, or trigger unsafe actions.
Common attacks include prompt injection, data poisoning, model theft, and abuse of AI-enabled workflows through over-privileged integrations. Defenders reduce the attack surface by limiting access, validating inputs, isolating models from sensitive systems, logging model activity, and testing for abuse cases before deployment. In vendor environments, the AI attack surface also includes the provider’s controls, since a weak third-party integration can become an entry point into the business.