Mitigations are the protections added around a model to reduce harm. Some are built into the model through training, such as teaching it to refuse dangerous requests. Others sit around it: filters and classifiers that check inputs and outputs, monitoring and rate limits, restricted tools and permissions, and sandboxes that isolate what the model can reach, such as blocking network access.
They matter in safety and security testing. A result such as how often an attack succeeded depends heavily on which mitigations were switched on. Tests are sometimes run with them removed, to see what the raw model can do, and again with them in place, to see what a real user would face.
No mitigation is perfect. Filters can be bypassed, and models can be talked into ignoring instructions, so the protections enforced outside the model in code and permissions are the most dependable.
When you read test results, check which mitigations were active, and when you deploy a model, layer several of them.