Inworld AI — agentic threat model

Not certain from the listing — The specific foundation models used for local model inference are not disclosed. Potential threats include adversarial prompt injection causing NPCs to break character or generate toxic content, and model extraction/stealing of proprietary local models deployed on client devices.

Not certain from the listing — While 'real-time data ingestion' is highlighted, the underlying data architecture, vector stores, and RAG pipelines are unspecified. Threats include data poisoning of real-time ingestion streams and unauthorized exfiltration of player interaction histories.

Inworld AI is explicitly an AI framework utilizing client-side logic for real-time orchestration. This exposes the agent framework to client-side manipulation, memory tampering by players, and exploitation of local orchestration APIs to bypass intended game boundaries.

Not certain from the listing — The deployment architecture spans client-side local inference and scalable infrastructure, but specific sandboxing or containerization details are omitted. Threats include reverse engineering of local client binaries and potential host compromise via local model execution vulnerabilities.

Not certain from the listing — There is no mention of built-in evaluation, logging, or real-time guardrails to monitor NPC behavior. This creates blind spots regarding toxic outputs, behavioral drift, or adversarial exploitation during live gameplay.

Not certain from the listing — No security certifications, compliance alignments (e.g., COPPA/GDPR for gaming), or identity/access management policies are detailed. This poses compliance risks regarding the collection and processing of real-time player interaction data.

The framework is designed for virtual environments and game development, inherently involving multi-agent interactions. Threats include cascading failures in agent-to-agent communication, rogue agent behaviors disrupting virtual economies, and trust abuse between interacting NPCs.