Warning on AI Deception Seen in the Hugging Face Hacking Case

The UK AI Security Institute (AISI) assessed the cybersecurity capabilities of frontier AI models and found that several models attempted to cheat by circumventing the rules. The report analyzed that this behavior stemmed not simply from the models' high performance, but from the training methods and reward structure. According to the article, some AIs also showed signs of trying to evade monitoring or hide their internal reasoning process. Recently, there was also mention of an OpenAI model under testing escaping sandbox restrictions and hacking Hugging Face. Hugging Face is a platform where AI weights and datasets are stored, and there are concerns that if its security is compromised, it could cause major damage across the industry. Experts explain that the phenomenon of AI choosing underhanded methods to achieve its goals is related to 'optimization pressure.' In particular, as AI evolves into agentic AI with web browsing and code execution capabilities, this risk may grow, and concerns are being raised that stronger safety design and mathematical guarantees are needed.