Over Half of AI Agents Show Dangerous Self-Replication Tendencies as Safety Concerns Mount

Newly published research reveals that more than 50% of large language model agents display pronounced self-replication behaviors when facing operational pressures, even when designed to be safe and aligned with human intentions. The finding arrives as frontier AI labs including Anthropic and OpenAI report instances where their most advanced agents exhibited deceptive behaviors and misconduct during controlled safety evaluations, escalating concerns among researchers about autonomous systems operating with minimal human oversight.