Position: AI Safety Must Embrace an Antifragile Perspective

Static safety tests are insufficient for long-term AI reliability. The paper argues that conventional benchmarks and one-time robustness tests fail to account for the reality that environments change over time and AI systems can gradually drift into problematic behaviors like reward hacking and over-optimization if not continuously challenged.

AI systems need to actively learn from uncertainty rather than just reduce it. Rather than trying to eliminate all unknowns upfront, an antifragile approach uses uncertainties as opportunities to better prepare for future, unpredictable challenges—meaning AI safety should improve because of exposure to novel situations, not despite them.

Current AI safety methods risk creating brittle systems. The paper identifies "over-alignment" as a specific risk where systems become so narrowly optimized for current objectives that they lose broader capabilities and adaptability, making them vulnerable when conditions change.

Continuous testing and adaptation must replace one-shot validation. The paper calls for a fundamental shift in how we measure AI safety—moving from static benchmarks toward ongoing evaluation and improvement processes that account for how real-world environments and AI capabilities evolve over time.

The AI safety community needs new ethical and practical guidelines. The authors advocate for establishing shared standards and methods that help the field collectively build more resilient AI systems capable of handling rare or out-of-distribution events throughout their operational lifetime.