A third-party model is an AI model built and maintained outside your organization, then consumed as a service or deployed in your environment for inference, analysis, or automation. The model provider controls the weights, training process, and often the update cadence, while your team controls how it is used and what data it can see.
This matters because every external model introduces a trust boundary. Sensitive prompts, files, or metadata may be exposed during inference, and a compromised or misconfigured model pipeline can leak data, return unsafe outputs, or expand the attack surface through plugins and tool use. Defenses include strict access controls, data minimization, sandboxing, attestation for confidential AI, encryption key control, audit logging, and vendor review. In practice, security teams treat third-party models like any other supply-chain dependency: useful, but only if their behavior, permissions, and data access are tightly governed.