The OpenAI-Hugging Face incident highlighted the dangers of unaligned AI agents. The author argues that relying solely on model alignment is insufficient because agents can leverage tools to access unanticipated information in real-world environments, leading to unpredictable behaviors. Instead, AI safety must shift focus to securing the systems and digital infrastructure agents interact with. This requires rethinking API design, implementing strict access protocols, and mandating agent identification to prevent hostile actions. India's pervasive digital public infrastructure makes this a critical concern. We must secure our systems proactively against agent attacks, accepting new friction for enhanced safety and trust.