LM Studio — agentic threat model

LM Studio directly loads and runs foundation models (LLama, Falcon, MPT, etc.) locally. The primary threats at this layer are adversarial prompt injection, downloading backdoored or malicious models from Hugging Face, and generating misaligned or harmful outputs locally.

Not certain from the listing — The description does not mention built-in RAG, vector stores, or data pipelines. The primary data risk is the potential exposure of local chat history files or the ingestion of poisoned local documents if external tools are manually connected.

Not certain from the listing — LM Studio serves as a model provider rather than an orchestrator. It does not appear to have built-in agent planning, memory, or tool-calling frameworks, though it exposes an OpenAI-compatible API that external agent frameworks can exploit.

The application runs locally on user hardware and hosts a local API server. Key threats include unauthorized local network access to the API port, potential remote code execution (RCE) via vulnerabilities in the model parser (e.g., GGUF format exploits), and lack of sandboxing for executed models.

Not certain from the listing — There is no mention of built-in guardrails, output filtering, logging, or evaluation metrics. Observability is limited to the user-friendly chat interface and local console logs.

The application provides strong privacy compliance by running entirely offline, preventing data exfiltration to third parties. However, it lacks enterprise security controls such as built-in authentication/authorization for its local API server.

Not certain from the listing — The app does not feature an active multi-agent ecosystem, but it connects to Hugging Face for model discovery, introducing supply chain risks from unverified third-party model publishers.