Human Supervision as Attractor Stabilization: Indirect Prompt Injection, Longitudinal Object Custody, and the HibriMind Ontology

Recent work on long-term memory poisoning, agent drift, trajectory-conditioned safety, shared editable states, and human-AI co-inference increasingly confirms a problem already formalized within the HibriMind corpus: in longitudinal human-AI interaction, the object of work is not merely transmitted through prompts and answers. It is progressively stabilized, displaced, contaminated, or transformed along the trajectory. This paper argues that human supervision is not only an error-correction mechanism, but a jurisdictional function that guarantees stabilization of the legitimate attractor organizing the interaction. Within the HibriMind ontology, the core risk of indirect prompt injection is not limited to malicious instructions embedded in external content. The deeper risk is the conversion of external content into an unauthorized operational attractor capable of redefining the trajectory while appearing locally useful or coherent. The paper distinguishes O₀, the authorized initial object or purpose; the operational attractor, which organizes local execution; and marginal attractors, which may arise from external sources, memory contamination, relational drift, or internal recomposition. It concludes that current agent-safety work remains largely situated in a human-tool paradigm, while HibriMind frames the problem as a genealogical human-algorithmic trajectory requiring object custody, provenance, reversibility, and human jurisdiction.