An open source AI model is one where the underlying weights — the billions of numerical parameters that define how the model behaves — are publicly released for anyone to download, inspect, modify, and run. Models like Meta's Llama are open source, meaning you don't need to access them through a company's API; you can run them on your own hardware.

The security debate centers on a fundamental tradeoff. Openness is great for transparency and research — anyone can audit the model, catch biases, or build on top of it. But it also means you can't "patch" a released model the way you'd patch software. If someone finds a way to misuse it — generating harmful content, for example — the weights are already out in the world and can't be taken back.

Critics argue open weights lower the barrier for bad actors to remove safety guardrails, since those guardrails are just additional fine-tuning that can be undone. Defenders counter that determined bad actors find workarounds anyway