If you`re developer - you should fix this
The same samples per pixel:
Out of curiosity- how many samples were each of these rendered with?
The anti-aliasing is due to the brightness of the light being reflected, since the color of the light would be a bit greater than 1 in this case, that would mean that the anti-aliasing filter would have a hard time trying to average the values into a smooth anti-aliased line, the solution would be to use a light path trick so that the light appears as somewhat dimmer in reflections while maintaining the brightness of the scene.
The bright points are called fireflies, they are seen in every pathtracer and represent caustics or other lighting effects that are undersampled, as a result they will average out over time, but how quickly depends on how small the light is (the smaller the light, the longer it will take). The solution to this (without using denoising techniques in a paint program or using a lot more samples) is something not implemented in Cycles yet, which is bidirectional path tracing, it differs from path tracing as it excels at rendering light from small sources.
they are seen in every pathtracer and represent caustics or other lighting effects that are undersampled
I know, but there are some algorithms of optimisation. Look at Maxwell Render images.
Yes, MLT and bidirectional path tracing are helpful, that’s not a new concept to anyone. They aren’t in Cycles atm, because Cycles was developed with a focus on being more an animation and motion graphics renderer than a totally realistic physics sim. As a result, it’s faster and has a few cheats to make it even faster still, like the “no caustics” options and the path depths per light type.
If you like Maxwell better, use it instead.