Please do not open a public issue for a vulnerability that could expose user data, execute unintended code, bypass a trust boundary, or expose a local inference service.

Use GitHub's private vulnerability reporting / Security Advisory workflow if the repository presents that option. Repository documentation does not assert that the feature is enabled. Otherwise, use a private contact method published on the maintainer's GitHub profile and initially share only enough information to establish contact.

A useful report includes:

• affected LLMRig version or commit
• operating system and Python version
• local inference backend/version
• reproduction steps
• expected and observed behavior
• impact
• a suggested fix, if you have one

Live discovery is metadata retrieval, not installation trust. A discovered repository remains informational until its runtime identifier and provenance are reviewed and added to the curated catalog. Users remain responsible for the trust, license, and content of models and runtimes they install.

LLMRig invokes local runtime processes and connects to a configurable Ollama service. The default is loopback. Do not expose an unauthenticated runtime service to an untrusted network, and treat a non-loopback --host or OLLAMA_HOST as an explicit trust decision.

Native GGUF and MLX-LM execution accepts explicit local paths. Retained private locator state is confined to non-serializable InventoryTarget / ExecutionTarget objects and passed to runtimes as argument-list values; public solve JSON, human output, and sanitized execution failures omit it. A user-supplied file association and structural check do not independently attest the artifact's identity, weights, safety, or runtime behavior.

Benchmark passports and public solve results are designed to exclude known private paths and machine identities, but review any output before sharing it. Hardware facts, model identifiers, runtime versions, and measurements may still be identifying in a particular context. Errors crossing public boundaries are categorical and sanitized.

Passport SHA-256 values provide deterministic identity and integrity checks. They are not signatures, cryptographic attestations of model contents, independent benchmark verification, or proof that a reported measurement is true. Path-independent native artifact identifiers can refer to different files on different machines.

Until LLMRig reaches a stable 1.0 release, security fixes are applied to the latest release line. Users should update to the newest available version before reporting an issue that may already be fixed.