“Closed, proprietary AI models are not inherently safe or secure, as they can be jailbroken, leaked, or have poorly engineered guardrails.”