What steps are typically part of a safety risk assessment for AI systems?

• A. Publish all model details publicly
• B. Identify stakeholders, assets, and threat models
• C. Rely on intuition alone
• D. Ignore potential harms

Answer: (B)

Systematic risk assessment for AI systems starts by mapping who will be affected, what needs protection, and how things could go wrong. The best approach is to identify stakeholders, assets, and threat models. Stakeholders are the people or groups with an interest in the system or who could be affected by its behavior. Assets include data, trained models, code, infrastructure, and the safety and privacy properties the system is expected to uphold. Threat models consider potential adversaries, failure modes, and ways those weaknesses could be exploited or lead to harm. Building this picture lets you evaluate risk, prioritize controls, and design effective mitigations like governance choices, access controls, monitoring, testing, and incident response. Publishing all model details publicly is not typical and can create misuse risks; relying on intuition alone misses concrete threats and gaps; and ignoring potential harms undermines the purpose of the assessment.