An open-weight model is one whose learned parameters, or weights, are released publicly for others to inspect, run, and sometimes adapt. This is not the same as fully open source: the weights may be available while the code, training data, or license terms remain restricted.
In cybersecurity, open-weight matters because it changes who can operate an AI system and how. Defenders may self-host the model to keep sensitive prompts, logs, and documents inside their own environment, and they can audit or fine-tune it for security tasks. The same availability also helps attackers, who can run the model without vendor controls, modify it for phishing, malware support, or other abuse, and deploy it behind weak safeguards. Whether open-weight is a security benefit or risk depends on the license, governance, and the controls around deployment.