* Results from the untrusted part inside container are returned using pickle. Which can be used to achieve arbitrary code execution outside container.
* no time limiting
* no memory limiting
* Untrusted code is run as root in the container which by default is same user as root outside container. From what I understand it isn't as bad as it was in earlier docker versions but still not great.
* untrusted code is run in the same process as semitrusted run_lang code, which means that the untrusted code with little bit of effort can manipulate reported execution time and memory usage
* for some languages correct_output is copied into the untrusted execution environment which means that solution could potentially just read the correct answers instead of calculating themselves
* none of the default capabilities are dropped which is probably more than what solution needs
* solution can freely create new files