Which term is used for the practice of deliberately testing to reveal weaknesses by adversarial means?

• A. Red-Teaming
• B. Adversarial Robustness
• C. Frontier Model
• D. AI Safety

Answer: (A)

Red-teaming is the practice of deliberately testing to reveal weaknesses by adversarial means. It involves simulating an attacker or adversary to probe a system, model, or process and uncover vulnerabilities, edge cases, or policy gaps so they can be fixed before real harm occurs. In AI safety, red-teaming helps surface prompts, inputs, or scenarios that cause a model to fail safely, leak sensitive information, or bypass safeguards, guiding improvements to defenses and training.

Adversarial robustness, in contrast, focuses on making a model more resistant to adversarial inputs, rather than the proactive probing and vulnerability disclosure that red-teaming emphasizes. Frontier model refers to an exceptionally capable model, not the testing approach, and AI safety is the broader field concerned with preventing harm, not the specific technique of adversarial testing.