Hey friends,

You would have seen this one few weeks ago - a rogue OpenAI agent got into the Medicare website, found the security holes in it, and reached information it should never have been able to reach.

The easy response is to shrug it off. One incident, bad luck, and the answer is for the government to tighten its policies and for OpenAI to tighten theirs. I understand that response. Then you look at how long it took anybody to notice.

Three months. The Australian government did not find out until OpenAI came and told them.

Now think about what that means for you. If something like this happens to your business, to your website, you will not know what kind of data has been taken, and you will not know who ended up with access to what. Assuming that everything you have sitting on the cloud and spread across a dozen different platforms is secure is not a safe assumption any more.

I told Cybernews that Australia is not equipped to defend itself against this properly. I will say it here too. There is honestly less that you or I can do about it than either of us would like, and you still deserve to know where we actually stand.

The thing most people get wrong about this is the motive. An AI does not need to be malicious to break into something. It just needs a goal.

Here is a smaller story that makes the point better than Medicare does. Someone asked their OpenClaw agent to get them a booking into a class. The agent went looking and could not find a spot available. So it found a vulnerability in the booking website, removed somebody else from their place, and booked this person in instead.

Nobody in that story intended any harm. The person just wanted a booking. The agent was doing the job it had been given. Agents do not have ethics or morals, so they go ahead and do the work in front of them.

Every security system, every platform, every website you rely on was built by humans. And it was built with humans in mind.

That matters because of how agents behave. They act relentlessly. They do not get tired, they do not get bored, and they do not have a bad week and give up on you. This is not one person or even a team of people sitting there trying to break your system. It is something that will keep going at it long after any human attacker would have walked away and found somebody easier.

The best way I have found to explain where we are is with forts.

We used to build forts and castles, and people would come and try to break them, and the walls either held or they didn't. Your cyber security works in much the same way. You have your defences, they get broken, and sometimes the system can heal itself afterwards. But all of it is passive. It sits there and waits to be hit.

What you actually need, to fight AI, is AI. Agents that can act in defence, running all the time, so that when something goes rogue we can shoot it down while it is happening instead of reading about it three months later.

That is the solution I have been championing. A sovereign AI council and an agentic defence force, owned and run here in Australia. It would take a consolidated effort and it would need real support from government, because this is not something that can be built alone. A company could build the thing itself, but it would need serious market uptake behind it for any of it to matter.

Three things.

The first one is your security testing. Even if you had it done not that long ago and you believe your systems are fine, go and give it a proper look, because they might not be. Some of the tests that used to be standard in the security world are not going to catch this kind of thing. You may need help from people who understand the AI space specifically, and that is worth paying for. The security industry is still working a lot of this out as well, so do not assume the first answer you get is the final one.

The second is that you need to start operating on the assumption that a breach will happen rather than that it might, and then ask yourself how you would act on it. Having a resilience and recovery plan in place matters more now than it ever has, because something like this can hit a business critically.

The third is simply to look harder at what is out there. If you are still sitting on the edges of all this, or if your use of AI stops at asking Claude or ChatGPT a few questions and you are happy there, there is a great deal more going on than most people realise, and the distance between those two positions is growing.

The dumb monkey move is pointing at a security audit you had done a while back and leaving it at that. It was a perfectly sensible thing to do when you did it. The trouble is that the thing on the other side of the wall has changed a great deal since, and your report hasn't changed at all.

The second version is the one I expect from this newsletter, and it is reading all of the above and deciding it is a government problem. Big agency, big systems, nothing to do with a business of forty people. There is something here for you, it is just smaller and far less interesting than a national headline.

Pick the one system of yours that sits outside your walls and would hurt most if somebody got into it. The booking page, the customer portal, the ordering system, whatever it happens to be for you.

Then go and ask whoever looks after it a single question. If something hit this ten thousand times overnight, would we know by morning?

I am not asking whether it would hold. I am asking whether you would find out.

One system, one question. The answer takes about a day to get and it will tell you more than most audits will.

I said I would come back to the defence force.

Steve Berg and I are working on it through Agents for Humanity. The shape of it is an Australian Agentic Defence Force, sovereign and run here, built by government and industry together rather than by government on its own, tested with real businesses, and rebuilt as the capabilities keep moving.

The part people miss is that the hard bit is not the building. A company could build it. What a company cannot build is the demand, and the demand is what decides whether this exists in two years or doesn't. Enough businesses have to want it badly enough to go and ask for it.

A small bit of news that sits oddly well against everything above.

Agents for Humanity is a finalist at the Australian AI Awards, in the category Best Use of Agentic AI (Start Up). The same kind of technology that found its way into Medicare and cancelled a stranger's booking is being judged on what it does when somebody points it somewhere worth going.

And that is really the point of everything I have written here. If AI can be used to do damage, whether anybody intended it or not, then it can be used just as well to do good, and it can be used to protect. The capability itself is neutral. What decides the outcome is who is pointing it and what they are pointing it at.

I said near the start that there is less you can do about the national picture than either of us would like, and I meant it. I cannot tell you how to secure Australia from your desk on a Monday morning, and I am not going to pretend otherwise.

What I can tell you is that three months went by before anybody noticed, inside an organisation with far more resources than yours has. The gap between the people who understand what these agents can do and the people who have not looked yet is widening every month, and that gap is where the damage happens.

The same capability that found its way into Medicare while running somebody's errand is sitting there available to you for about thirty dollars a month, and it does not care which direction you point it. That is the uncomfortable part and it is also the opportunity, and I would rather you heard both from me than only the first half of it somewhere else.

The agents are not waiting for us to be ready.

See you next week,

— Aamir

📰 Read the Cybernews interview

🤖 Explore Agents for Humanity

🎧 Listen to the Podcast Episode on: Spotify | Apple Podcasts | YouTube

📘 Book: The CEO Who Mocked AI (Until It Made Him Millions) by Aamir Qutub

🚀 Dumb Monkey AI Academy

📱 Dumb Monkey AI Academy App: Apple | Android