This question came up from a tangential discussion. Some one was asking suggestions about systems architecture. I suggested them to look into RCA (root cause analysis) for some big scale failures. Hard part is to find this kind of data publicly.

This prompted the question about the data source for Mythos/Fable training. The security bugs reports, fixes are somewhat of a sensitive private collections.

I did check the documentation from Anthropic on this, it only mentions the usage of publicly accessible data.

Any ideas how Anthropic would have got their hands on this private data if they used that at all?

It was trained from experts telling older models what to do. They had agreements with these people who already did this stuff all day.

The benefit with the models is that you can get rid of alot of the grunt work and just try a massive amount of things that might work all at once in a short amount of time in a automated fashion.