Don't use bearer tokens for security, and don't give your AI access to anything you don't want it sharing. It's that simple.

Well then generating some test coverage sounds like a good play.

The hyperscalers are all renting capacity from each other. This sort of thing allows them to run frontier models without the owner of the hardware spying the weights.

That will be $10000000 please

I would say it's mostly a matter of understanding what it can do, which keeps changing, and how best to use it.

In many ways, AI is forcing people used to doing their own work to learn how to delegate. That's also a skill.

AI, at least the frontier models, should be really good at this. Ask it to evaluate test coverage and generate unit tests for any missing behaviors. Ask it why this function does this thing. Ask it to generate the exact same behavior in a different language. It will do all of that well, fast, and probably catch things a human would miss.

It isn't perfect. It isn't replacing you. But it will help you do your job better.

The skill is in knowing whether or not what it did is the right thing, or requires refinement. Same skill as any code review.

You do most of the design work. You make decisions which the AI is good at calling out. What you don't do, is take an hour to go from concept to code. You write the concept, hit return, and move on to something else. You can juggle 5 or 6 tasks effectively this way.

If there's 900 privilege escalation attacks on ANY version something is wrong. And priorities need to be re-evaluated in the AI era.

948 paths to admin, and only 10 of them were unknown? What are you doing, Linux?

I'm not sure what model backs the AI overviews but it's probably a 3-series flash model. Gemini 4 Argon is pretty good.

Google' AI Mode tab, or the Gemini app, are both much better than AI overviews.

Not completely. Gemini is pretty good. Their mistake was not recognizing the opportunity to go from research to product.

It's a coding "assistant" in the sense that it wants to do all the coding. It assumes everything you need to tell it can be done in natural language.

The chain of thought output sure does LOOK like reasoning. The end results look as they would if it were reasoning. We're getting into "a difference that makes no difference is no difference" territory here.