As much as I love Python and its ecosystem (look no further than my Medium and GitHub profiles), I have to disagree with you here -

None of those things utilize Python at their core - the language merely serves as a wrapper for the services provided by these packages as an API (for good reason since readability and accessibility is what Python does best).

PyTorch, Tensorflow, Theano, Scikit (both depend on numpy), and almost any other popular library you can find "for Python" is not actually fully implemented in Python - the performance critical parts are almost always siphoned off to C or C++ using interpreter extensions.

Code written for penetration testing purposes is rarely required to be "fast" in the most traditional of senses.

The popularity of langauge has nothing to do with its exeecution speed; it is entirely irrelevant.

The bottleneck aspect you talk about rises quite quickly in large scale industrial systems and you can be guaranteed to suffer at the hands of Python at some point or the other. Instagram itself has tried to squeeze water from stones using Django as their backend but ultimately had to resolve to using hacks such as disabling the GC for the CPython interpreter including numerous other changes that their Python codebase barely resembles the Python we're all familiar with.

As an ending note, I encourage you to check out the Computer Language Benchmarks Game for some truly eye opening differences in performance where it truly matters - and one cannot even complain of poor code since the implementation of all the "games" is almost entirely the same using similar language semantics and features.