Before and after releasing an AI system, developers run red-teaming: people and tools deliberately try to make it behave badly, for example producing harmful content or revealing information it should not. The weaknesses found are then used to strengthen the system.
This adversarial testing is now a standard part of responsible development, treating an AI model like security software that must be attacked to be trusted. It cannot find every flaw, and new weaknesses emerge as systems and their uses change, so red-teaming is ongoing rather than a one-time check, pairing human creativity with automated probes.