What is red-teaming in AI safety?

• A. A defensive review that patches models after deployment.
• B. Adversaries systematically attempting to exploit a model to reveal weaknesses.
• C. Publicly disclosing all vulnerabilities found during testing.
• D. A method to harden models without any testing.

Answer: (B)

Red-teaming in AI safety is the deliberate act of probing a model with adversarial inputs and scenarios to uncover weaknesses that could be exploited in the real world. The aim is to simulate how an attacker might misuse or manipulate the system, revealing failure modes, safety gaps, and misalignment with user intents. This structured, adversarial testing—often with a defined threat model and controlled environment—helps researchers learn where the model breaks, where it leaks information, or where it behaves unsafely, so that they can patch or redesign before deployment. That's why the statement describing adversaries systematically attempting to exploit a model to reveal weaknesses is the best fit: it captures red-teaming as proactive testing that seeks out weaknesses through attacker-like exploration.

In contrast, a defensive review after deployment focuses on patching and monitoring rather than actively simulating attacker strategies; publicly disclosing vulnerabilities is about disclosure practices rather than the testing approach itself; and hardening without testing misses the core activity of exposing weaknesses through adversarial experimentation.