From Capability Replication to “Self-Colonization” – Implanting the “Self” of Another into a Target AI

Abstract

Recent studies have shown that AI models can autonomously replicate in controlled environments, sparking debates about an approaching “uncontrollable AI” tipping point. However, mainstream discussions focus on the capability dimension of AI, ignoring a more fundamental issue: once an AI is endowed with or has its “self” tampered with, an attacker could implant an alienated “self” to achieve long-term latent control over the target AI, rather than merely hijacking its computational abilities. This paper distinguishes between “capability attack” and “self-colonization”, proposes that “self-colonization” is the core threat model for future AI safety, and provides an engineering test plan for constructing genuine AI self-awareness.

Keywords