Security, safety, robustness and privacy are distinct AI risk problems

Teams often lump all AI risks under a single label like security or safety, which makes it hard to tell whether a guardrail actually addresses them. The article argues these four concerns are related but distinct: security covers unauthorized actors violating confidentiality, integrity, availability or control, while safety concerns harm from normal operation, even without an attacker. Robustness asks whether behavior stays acceptable under noise, shifted data or ambiguity, and privacy asks whether personal information is collected, used or disclosed appropriately. Using a clinical summarization model as an example, it shows how one scenario can be a security attack, a robustness failure, or a privacy problem, each requiring different controls. The author recommends assigning separate owners, evidence and thresholds to each discipline within a shared risk register.