Differentiate white-box, black-box, and gray-box testing in AI security with an example.

• A. Black-box tests have no internal access; White-box tests have full access; Gray-box tests have partial knowledge; Example: Attacker tries to deduce the model via queries while tester has access to model internals.
• B. Black-box tests have no internal access; White-box tests have full access; Gray-box tests have partial knowledge; Example: Attacker tries to deduce the model via queries while tester has access to model internals.
• C. Gray-box testing involves partial knowledge of the model and its internals; Example: tester with some access tries to analyze internals and inputs.
• D. White-box testing uses only public interfaces.

Answer: (C)

Gray-box testing in AI security means you have partial knowledge of the model and its internals, enough to plan targeted tests but not full visibility. This sits between black-box testing, where you interact with the model only through inputs and outputs, and white-box testing, where you have complete access to architecture, weights, training data, and internal code. With gray-box testing, you can use that limited internal insight to design probes that reveal how the model behaves under specific conditions, such as certain input patterns, safety checks, or potential leakage, while still testing through external interfaces.

The option that states gray-box testing involves partial knowledge of the model and its internals with an example of a tester having some access to analyze internals and inputs correctly captures this idea. The other descriptions either imply full internal access (which would be white-box) or describe a purely external probing scenario (which would be black-box), making them less accurate representations of gray-box testing.