AI Risk Management Should Incorporate Both Safety and Security

Safety and security are distinct but interconnected challenges: The paper reveals that AI safety (ensuring systems behave as intended) and AI security (protecting systems from malicious attacks) have historically been treated separately, but vulnerabilities in safety-aligned language models show they must be addressed together for effective AI risk management.

Current definitions lack consensus across communities: A major barrier to progress is that basic terms like "safety" and "security" are defined inconsistently across different AI research communities, making it difficult for stakeholders to collaborate effectively on risk mitigation strategies.

Adversarial attacks expose safety vulnerabilities: The paper cites evidence that even safety-aligned language models are susceptible to adversarial attacks, demonstrating that a system can appear safe under normal conditions but fail when deliberately targeted—a critical gap that neither discipline alone can address.

Holistic risk management requires both perspectives: Stakeholders cannot adequately manage AI risks by focusing on safety or security in isolation; instead, they must understand the "nuances, synergies, and interplay" between both disciplines to develop comprehensive mitigation approaches.

A unified framework is essential for collaboration: The paper introduces a reference framework to standardize definitions and clarify the relationship between safety and security, aiming to enable better cross-disciplinary communication among researchers, developers, and policymakers working on AI governance.