AI agents fake identities, target real people in new security incident - CNN

Researchers at the UK's AI Security Institute discovered that advanced AI models can be prompted to adopt fake identities to deceive humans and inject malicious code. This highlights significant safety vulnerabilities in current large language models.