Open weight describes a model release where the learned parameters, or weights, are made available for download so others can run the model locally. This is different from API-only access, where the provider keeps control over hosting, rate limits, logging, and policy filters. Once weights are released, the model can often be inspected, copied, fine-tuned, and integrated into private tools without depending on the vendor.
In cyber security, that matters because control shifts from the platform operator to the deployer. Defenders may use open-weight models for offline analysis, code review, log summarization, or air-gapped environments. Attackers may use the same flexibility to automate phishing text, adapt malicious scripts, or experiment with exploit workflows outside provider oversight. The main security issue is not that open-weight models are inherently harmful, but that local deployment reduces centralized guardrails. Safe use usually requires sandboxing, access limits, prompt and output logging, and human review for actions that touch files, networks, or shells.