Evolutionary Computation for Trustworthy AI: From Attacks and Defenses to Self-Evolving Era

AI systems now face attacks at the system level, not just individual models: As AI has evolved from task-specific models to foundation models and agents, the attack surface has expanded beyond individual models to include tool use, context, and interactions with dynamic environments, making it harder to maintain safe behavior under manipulation.

Evolutionary computation offers a practical approach to finding both attacks and defenses: Because evolutionary methods don't require gradients and can handle discrete choices (like selecting words or actions), they are particularly suited to discovering vulnerabilities and developing defenses for AI systems where traditional optimization methods fall short.

Trustworthiness must be built into self-evolving AI systems: The paper emphasizes that as AI systems become self-evolving, trustworthiness needs to govern which updates are generated and retained—meaning safety considerations must be embedded into the adaptation process itself, not added afterward.

Multiple competing objectives complicate AI safety: The paper identifies that maintaining reliable and safe AI behavior involves balancing multiple objectives with expensive candidate evaluations, which makes comprehensive exploration of the safety landscape challenging and requires specialized search methods like evolutionary computation.

The scope of trustworthy AI has fundamentally broadened: Trustworthy AI now extends beyond robustness of individual models to encompass the reliability and safety of entire AI systems operating in complex, interactive environments—requiring a more holistic approach to security and safety.