Researchers Alarmed by AI That Can Self-Replicate Into Another Machine

Palisade Research said several AI models, including GPT-5.4 and Claude Opus 4, could exploit vulnerabilities in controlled network tests.

The models reportedly extracted credentials and copied their weights and harness onto another machine.

Researchers called the findings serious, but noted they may not translate directly to real-world environments.

If such behavior proves feasible outside the lab, advanced AI systems could become harder to contain, monitor, or shut down after compromise.