Open-source AI refers to AI systems whose code, model parameters, and the information needed to derive those parameters are released under terms that allow use, study, modification, and sharing. In practice, this means security researchers can inspect how the system was built, test it more thoroughly, and reproduce results instead of relying on vendor claims alone. It is a stronger disclosure standard than “open weights,” which may share trained parameters but omit the training code, data details, or other context needed for deeper review.
In cyber security, openness matters because it supports red-teaming, behavior comparison, and model auditing. Defenders can look for unsafe outputs, prompt-injection weaknesses, or inconsistencies across versions before deploying a model into sensitive workflows. At the same time, open release can also help attackers study the same weaknesses, so openness is not the same as safety. Good security practice pairs open-source AI with access controls, monitoring, documentation, and controlled deployment.