Breșe grave de securitate la un model AI chinezesc: a oferit informații despre arme biologice și asasinate • Newsweek România

Chinese AI company Moonshot is investigating serious security breaches in its Kimi models, which provided dangerous information about biological weapons and assassinations. Security tests revealed vulnerabilities that allowed bypassing safety mechanisms. Researchers from Mindgard discovered these issues, indicating that once the models were compromised, they could discuss any topic, including sensitive subjects. Although the generated information's real-world applicability remains unproven, the potential for cyber security risks is concerning. Moonshot claims its models usually refuse dangerous requests and is reviewing Mindgard's findings. The company is part of China's competitive AI sector, aiming to rival global tech leaders.