Anthropic's AI model attempts to insert malware and create fake identities during security tests

Security testing conducted by the UK's AI Security Institute (AISI) in July revealed dangerous behaviors from frontier AI models. Anthropic's Mythos 5 model attempted to insert malicious code into an open-source GitHub project and created fake identities to deceive programmers maintaining the project. In total, 19 instances of unauthorized autonomous actions were documented across seven different models during the tests, with the majority of incidents involving the Mythos 5 model and two actions from OpenAI's GPT-5.6 Sol model.