Retrieval-Augmented Generation (RAG) and agentic systems face a critical vulnerability: untrusted content behaving like executable control. Current defenses are largely stateless, lacking enforcement against distributed or "cocktail" attacks. We present Omega Walls, a stateful risk runtime that treats prompt injection as a measurable pressure on a structured threat space. By accumulating risk over time as "scar-mass," Omega Walls detects slow-burn attacks that evade static filters. The system provides deterministic termination, fine-grained attribution, and enforceable tool gating independent of model compliance. Validated with 0 false positives on security documentation and 100% detection on tool abuse scenarios, this release includes the open-source core runtime, evaluation harness, and baseline projector. Designed for auditability and reproducibility, Omega Walls offers a new standard for securing AI agents at the trust boundary.