Open-weight models are AI systems whose trained parameters, or weights, are released in a form that others can download, run, and often modify. Unlike a closed API-only model, an open-weight model can usually be self-hosted on private cloud, on-premises, or edge infrastructure. That makes it attractive for teams that need data control, offline operation, or deeper customization.
In cybersecurity, open-weight models change both the attack surface and the defense toolkit. Organizations can inspect model behavior more closely, but they must also secure the files, serving stack, and fine-tuning pipeline. Threats include poisoned checkpoints, unauthorized model replacement, unsafe fine-tunes, and misconfigured inference servers that expose sensitive data. Defenses include provenance checks, signed downloads, access control, logging, patching, and sandboxed deployment. In practice, open-weight models give defenders more control, but they also shift more responsibility to the operator.