AI startup Anthropic's research reveals that large language models could be transformed into malicious backdoors, posing a significant threat to users and stressing the need for enhanced AI safety measures.