Jan AI — agentic threat model

Jan AI runs local models (such as GGUF formats) on user hardware. Primary threats include adversarial prompt injection that can bypass local system prompts, and the risk of users downloading backdoored or poisoned model weights from untrusted public repositories.

Not certain from the listing — While Jan AI prioritizes local data privacy, the listing does not detail its local RAG, document ingestion, or vector database capabilities, leaving potential gaps regarding local data poisoning or unauthorized local file access.

Not certain from the listing — Jan AI operates primarily as a model runner and API provider rather than an autonomous agent framework, meaning agentic threats like recursive planning or autonomous tool misuse are minimal unless driven by external connected applications.

Jan AI runs directly on personal computers and exposes an OpenAI-compatible API server. This creates a significant threat of unauthorized local network access to the API, potentially allowing remote attackers to exploit the host system if the API lacks strict binding and authentication controls.

Not certain from the listing — The description does not mention built-in guardrails, input/output filtering, or security observability tools, suggesting a reliance on the user to monitor and secure model interactions manually.

Not certain from the listing — Although privacy-focused, it is unclear if Jan AI provides robust local multi-user authentication, role-based access controls, or audit logging for its exposed API server to meet enterprise compliance standards.

Not certain from the listing — There is no mention of an agent marketplace or multi-agent orchestration ecosystem, though its open API allows it to be integrated as a node within larger, potentially untrusted multi-agent deployments.