Red Teaming Contemporary AI Models: Insights from Spanish and Basque Perspectives

All tested AI models showed significant safety vulnerabilities: Across 670 conversations, biased or unsafe responses were found in all three models tested, ranging from 29.5% in OpenAI's o3-mini to 50.6% in Spain's ALIA Salamandra, demonstrating that safety challenges persist even in recently developed systems.

Smaller, locally-developed models may face greater safety challenges: ALIA Salamandra, Spain's public AI infrastructure designed to support Spanish and Basque languages, showed the highest rate of problematic responses (50.6%), suggesting that smaller language models may require additional safety work compared to larger commercial alternatives.

Red teaming by human experts remains essential for finding AI vulnerabilities: The study used ten participants to manually test models through creative adversarial approaches, uncovering safety issues that standard testing may miss—highlighting the importance of human expertise in responsible AI development.

Language-specific AI systems need tailored safety testing: The focus on Spanish and Basque language models reveals that AI safety concerns aren't universal; systems designed for specific languages and regions require dedicated evaluation to identify culturally-relevant biases and safety gaps.

Global AI competition creates pressure that may compromise safety: The paper frames AI development within a competitive landscape between the U.S., China, and Europe, suggesting that the race for AI leadership could incentivize faster deployment over thorough safety testing and bias mitigation.