I think at least part of the problem is that the script is setuid, and there is/was a race condition in kernels when they execute a script and then launch the (non-setuid) interpreter for (what the kernel believes to be) that script, which in turn tries again to load the script. If you are/were careful to introduce a high CPU load and launched the whole conoction through a symlink, you could readdress the symlink to point somewhere of your choosing in the time between when the kernel launched the interpreter and when the interpreter read the script to process.

Note that all my experience here is from hearsay - I've never looked into the real mechanics of various kernels.