What is a key insight regarding security considerations in AI?

• A. Security must be applied at previously unconsidered points, like securing training data.
• B. Security can be added after deployment.
• C. Security only affects software, not data.
• D. Security is optional for AI.

Answer: (A)

Security for AI must be integrated across the entire lifecycle, starting with how data is collected, stored, and used. The key insight is that training data drives how the model behaves, so protecting that data is essential. If training data can be tampered with, mislabeled, or otherwise compromised, the model can learn the wrong patterns, behave unpredictably, or even include hidden backdoors. That’s why security needs to be applied at points that were previously overlooked—such as securing training data, ensuring data provenance and integrity, and guarding the data supply chain.

Putting security in only after deployment misses critical risk surfaces. Once a model is in use, attackers can exploit weaknesses related to data inputs, model extraction, or privacy leaks, which are harder to remediate. And it’s not correct to say security affects only software or that security is optional for AI—the data and the model’s training process are integral to AI safety and trust.