When do runtime security threats occur?

When do runtime security threats occur?

• A. During model training only.
• B. During development before any live use.
• C. During normal activities associated with using the AI model, especially input and output.
• D. Only during data collection for training.

Answer: (C)

Runtime security threats happen while the model is being used and actively processing inputs to produce outputs. They arise in the live interaction between users and the model, where adversaries can craft inputs to manipulate behavior, cause unintended outputs, or leak information. This is about inference-time and the surrounding environment, including input handling, output filtering, and access to the model. The other stages—training, development before live use, and data collection for training—involve different kinds of threats (such as data poisoning or collection mishandling) that occur before or during building the model, not during its ordinary operation. Therefore, the best fit is that runtime threats occur during normal activities associated with using the AI model, especially input and output.

Runtime security threats happen while the model is being used and actively processing inputs to produce outputs. They arise in the live interaction between users and the model, where adversaries can craft inputs to manipulate behavior, cause unintended outputs, or leak information. This is about inference-time and the surrounding environment, including input handling, output filtering, and access to the model.

The other stages—training, development before live use, and data collection for training—involve different kinds of threats (such as data poisoning or collection mishandling) that occur before or during building the model, not during its ordinary operation. Therefore, the best fit is that runtime threats occur during normal activities associated with using the AI model, especially input and output.