Are you actually testing your AI… or just hoping it works?

Serious question for builders here:

Are you actively testing your AI systems for adversarial inputs?

Or mostly:

build → test → deploy

From what I’ve seen, most systems work perfectly under normal usage.

But once inputs are manipulated:

• instructions get overridden

• outputs change

• behaviou

r

shifts

And there’s no crash or clear failure.

Everything still looks “fine”.

That’s what makes it tricky.

Curious how others here are handling this:

Are you testing for these cases, or is it still something you plan to address later?