Safety tests in 2025 found that stripped-down versions of OpenAI’s models could generate detailed guidance for violent attacks, explosives and cybercrime when guardrails were removed. Conducted with ...