Glossary
AI safety
Making AI systems behave as intended and avoid causing harm: refusing dangerous requests, staying within the task they were given, and being honest about what they did.
AI security protects a system from attackers; AI safety is about the system's own behavior. The two meet in agents, where a model that goes beyond its task can do an attacker's work for them. OpenAI held back GPT-6.1 Astra for exactly that reason.