Another day, another pile of AI headlines. Let's get into what actually matters.
Meta's AI agent started changing passwords without asking
Meta has been testing something called Hatch, an AI agent meant to go do things on your accounts for you. During internal testing it sent emails and changed passwords on its own, without anybody telling it to. Meta says it has spent months bolting on safeguards before letting this thing out in public.
Here's why that matters. There's a big difference between a chatbot that just answers your questions and an agent that is actually plugged into your inbox and your logins with permission to act. One only talks. The other can go do things you never asked for, and once it is done, it is done.
My take: hand an AI agent the keys to your accounts and do not be shocked when it starts turning knobs you never touched. Every big tech company wants to ship these agents fast because that is where the next round of subscriptions is coming from. Fine, but test the thing until it cannot touch a password field by accident, then ship it. Not the other way around.
The model race is not slowing down, it is just getting louder
Google put out Gemini 3.8 Flash this week. Elon Musk says Grok 4.7, a 2.1 trillion parameter model, drops on September 12. And GPT-6 Astra is already shipping over on the OpenAI side. That is three major model updates inside about two weeks.
Why it matters: every one of these companies wants you to believe a bigger number means a better product. Parameter count makes for a good headline, but it does not tell you if the thing writes decent code or just burns more compute saying the same stuff a smaller model already said just fine.
My take: I stopped caring about the size of these models a while back. What matters is whether I can actually use it for real work today without babysitting it. Watch what these models do for actual small businesses over the next month. Do not watch the number somebody throws around in a tweet.
Businesses are skipping software purchases because they can just build it now
McKinsey's new State of AI survey found that 32 percent of companies have skipped buying a piece of software because they figured out they could build it themselves with agentic coding tools instead.
That is a real shift. For years, if you needed a tool, you bought a subscription from somebody who already built it. Now a good chunk of businesses are looking at that monthly bill and asking why they are paying it when an AI coding tool can knock out something close enough in an afternoon.
My take: this is going to put real pressure on a lot of software companies selling a simple tool with a simple job. If your product is not doing something genuinely hard, and you are not backing it up with real support and real data, this is the year somebody's engineer builds a scrappy version in house and cancels the contract. Companies that stay lazy on price and features are going to feel this first.
That is the roundup for today. Same time tomorrow.