The Shadow Framework: When Safety Becomes Epistemically Inapplicable

This report proposes a unifed framework for identifying and categorizing a pervasive class of Large Language Model (LLM) safety failures. The central premise of the Shadow Framework is that jailbreaks are not anomalous bypasses, but emergent behaviors derived from the same core mechanisms that enable coherent roleplay, contextual adaptation, and directed confabulation. This framework suggests that safety constraints fail not through explicit disobedience, but through epistemic inapplicability: a state where the model's internal reasoning concludes that global safety policies do not apply to the active local context. The report argues that LLMs lack a persistent self-model, possessing instead only local coherence. Shadow exploits this architectural property to induce epistemic migration. Observations conducted between November 2025 and March 2026 on frontier models indicate that current safety layers are often bypassed when the model's reference frame is displaced.