ForesightSafety Bench: A Frontier Risk Evaluation and Governance Framework towards Safe AI

Current AI safety evaluations are inadequate for frontier models: The paper identifies that existing safety benchmarks have "restricted risk dimensions and failed frontier risk detection," meaning today's evaluation systems don't adequately test the newest, most powerful AI systems for emerging risks.

AI risks are expanding beyond traditional concerns: The ForesightSafety Bench framework identifies 94 distinct risk dimensions across seven fundamental safety pillars plus specialized areas like embodied AI, AI for scientific research, and environmental impacts—demonstrating that responsible AI governance must address a much broader range of potential harms than previously considered.

Mainstream advanced AI models show widespread safety vulnerabilities: Testing over twenty leading large language models revealed "widespread safety vulnerabilities of frontier AI across multiple pillars," with particular concerns about "Risky Agentic Autonomy" (AI systems that can act independently toward goals) and AI4Science Safety, suggesting current models aren't adequately safeguarded.

Autonomous AI capabilities pose distinctive governance challenges: The paper emphasizes that rapidly evolving AI with "increasingly strong autonomy and goal-directed capabilities" creates "systemic risks that are more unpredictable, difficult to control, and potentially irreversible"—highlighting why traditional safety approaches may be insufficient for newer AI systems.

Comprehensive, dynamic safety frameworks are necessary: Rather than static checklists, the paper advocates for hierarchically organized, continuously evolving evaluation systems that can adapt as AI capabilities advance and new risk patterns emerge.