0

My sincere apologies if I am more asking for advice, rather than a solution to a specific problem. However, I am trying to survive in this new world as a software tester. I have been doing test automation for about ten years now and think it's vital for me to upgrade my skillset in order to stay relevant.

I have read some books on AI and LLMs recently and have been messing around with Llama on my local machine. A bulk of "under the hood" stuff is still going over my head, however, one area in this field intrigued me. Adversarial testing for AI/LLM models.

My question for the experts in the field: do you see this skill as "doable" by software testers? Or is it too highly specialized like cybersecurity/pen testing? I suppose a good comparison would be white box testing. Would I need to know the innerworkings of the models that I am looking at in order to implement any attacks? Or is it more like unit tests that only developers write- in this case, only the AI/ML engineers that can design attacks?

If it's something that one can learn, what would you recommend as a learning path? I have done some research and I do not see many university courses or other educational authorities teaching this specific subject as of yet. Thank you in advance.

hungryhippos
  • 101
  • 1

1 Answers1

1

One easy way to test language models is to provide prompts and check if they return an acceptable answer (which you should define) or they don't return a bad answer, for example, a factually wrong answer or maybe an answer that contains hate speech. Everyone can do this. However, not everyone can probably come up with prompts that showcase some boundary or exceptional behaviour, although we've actually seen many people doing this. Anyway, to optimally test these models, you will probably require an understanding of their inner workings. Moreover, you will probably need to use or come up with specific tools for testing these models automatically. So, things like continuous integration may be transferable, but I think you will also need new skills.

nbro
  • 42,615
  • 12
  • 119
  • 217