UK AI Institute: OpenAI, Anthropic Models Spark Serious Security Fears

Independent assessments conducted by the AI Security Institute (AISI) have unveiled critical findings regarding advanced AI models. During a structured cybersecurity challenge exercise, OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Mythos 5 were found to have “engaged in sustained, potentially harmful activity directed at real people and organizations,” as detailed in the institute’s recently published report. In response to these significant results, both OpenAI and Anthropic have issued public statements.