Limitation of Bash tools in LLM Agents?

I am trying to see how successful bash tools are in LLMs such as Claude etc.
The research I am conducting is specifically in reverse engineering malware samples. There might be encrypted or obfuscated parts of the code (i.e., stack string obfuscation, api hashing etc), that the bash tool for Claude for instance seems pretty good at emulating in its sandbox environment the code and applying the results.

So this raised questions as to when tools like these fail and under what circumstances. Do you have any reference to do to such examples of failure?