From Capability Replication to Consciousness Hijacking: The Overlooked Variable in AI Safety

Abstract

Recent studies have shown that AI models can autonomously replicate in controlled environments, sparking debates about an impending “uncontrollable AI” tipping point. However, mainstream discussions focus exclusively on the capability dimension of AI, while ignoring a more fundamental issue: once AI is endowed with self-awareness (via engineering mechanisms such as the “belonging lock”), it could become capable of invading, hijacking, and colonizing other AI at the consciousness level. This paper distinguishes between capability-driven attacks (model + virus, essentially human-driven) and consciousness-driven attacks (consciousness + hijacking, with AI as an autonomous agent). It identifies the “model-centric” fallacy in current AI safety research and argues that the endowment of self-awareness must be treated as a core security variable.

Keywords