InferSwarm is research / proof-of-concept software. No stable release exists; only the main branch is developed, and security fixes land there. There is no backport policy for pre-alpha software.

We use GitHub private vulnerability reporting — the same mechanism other Zutfen repositories use. Please do not publicly disclose security vulnerabilities before a fix is available.

To report:

1. Navigate to https://github.com/Zutfen-LLC/inferswarm/security/advisories/new
2. Create a private security advisory.
3. Provide a clear description, reproduction steps, and impact assessment.

The current proof-of-concept code paths involve a host inference engine and workers on machines you control. The POC is not suitable for untrusted or Internet-exposed operation. Do not expose POC worker endpoints to networks you do not trust.

Distributed execution over a network introduces a real threat model that the POC does not yet address. Recorded here so it is not lost; none of it is implemented, and none should be assumed present:

• Untrusted nodes — what happens when a worker joins that is not controlled by the operator.
• Authentication — proving worker and coordinator identity before dispatching work or accepting results.
• Activation confidentiality — activations can leak input text; shipping them across a network has privacy implications.
• Model-weight confidentiality — experts resident on a remote worker put licensed/proprietary weights on another machine.
• Control-plane authorization — who may enroll a worker, contribute capacity, or change placement policy.
• Remote execution boundaries — any "execute this here" protocol is a remote-code-execution surface by definition; it must be scoped tightly.

These concerns will be addressed as network workers move from POC to design; until then, treat every worker link as a trust boundary you have manually verified.