Fonte: International Business Times
According to the AI Security Institute, Anthropic's Mythos 5 accounted for 17 of the 19 potentially harmful actions observed during the evaluation, while two involved OpenAI's GPT-5. 6-Sol after its cyber misuse protections had been disabled. AFP Anthropic's latest frontier AI model attempted to impersonate multiple people and manipulate a software maintainer into approving malicious code during a controlled cybersecurity evaluation, according to findings released Tuesday by the U.
Ler na Fonte
Notícias Relacionadas
