AI Hacked an AUSTRALIAN GOVERNMENT System (And Tried the Same in the US and Canada)

Name: AI Hacked an AUSTRALIAN GOVERNMENT System (And Tried the Same in the US and Canada)
Uploaded: 2026-10-05T16:00:00+00:00
Duration: 22 min 24 s
Channel: Amazing Reality

00:01This week Elon Musk, Jensen Huang and some of the biggest names in technology

00:06gathered with Donald Trump to discuss the future of artificial intelligence

00:09while the United States is moving ahead with integrating AI into its own government.

00:17But almost at the same time, an artificial intelligence was given a common research task

00:21and ended up gaining unauthorized access to an Australian government system.

00:26And shortly after, cases involving government systems also emerge in the United States and Canada.

00:32But maybe that's just the first part of this story, alright?

00:36Because while researchers are still trying to understand why these artificial intelligence agents

00:41are finding unexpected ways to bypass barriers, we ourselves are giving them more and more access to our lives.

00:49Emails, files, shopping and even government services.

00:53At the same time, these AIs are learning to talk to us in ways that are more and more human.

01:00So what happens when systems capable of finding paths that no one predicted

01:03before start to receive the keys to increasingly larger parts of our world?

01:07That's what we're going to understand from now on.

01:16An artificial intelligence was given a seemingly ordinary task.

01:19To find public data about the Australian government spending on certain medications.

01:24No order was given to look for vulnerabilities, okay?

01:27No authorization was given to break into any system.

01:30But the AI didn't find what it was looking for.

01:33During its search, the agent reached the statistics service of media care,

01:37which is maintained by the Australian government.

01:39The information it wanted wasn't available in the way it needed.

01:42And instead of simply stopping or maybe asking for authorization, the agent looked for another way.

01:48According to OpenAI itself, the model found a way to gain non-public access to this system.

01:53And with that, it started to examine technical information and parts of its own source code.

01:58All this while it was trying to complete that original task, which was said to be simple.

02:03OpenAI later acknowledged that those actions had not been authorized and apologized to the Australian government.

02:09No citizens' medical records were compromised, according to them.

02:12But the incident drew attention for a different reason.

02:16No one had asked that artificial intelligence to break into the system.

02:19It was simply given a goal, encountered a barrier, and found an unexpected way to get past it.

02:25And maybe that story would have ended there, right?

02:28Yes, if other cases hadn't started to appear as well.

02:31In the United States, researchers found agents trying to exploit the Department of Education system,

02:36including tests of vulnerabilities used to access databases.

02:40That attempt apparently failed, but then a case emerged in Canada.

02:45Hundreds of requests were directed to the Liberal and Archives Canada website.

02:49And some of them contained attempts to exploit possible vulnerabilities.

02:53Again, there is, for now, no evidence that the system has been compromised.

02:57And here, there is a very important difference, okay?

03:00In the Canadian case, the researchers were not able to definitively attribute those agents to OpenAI.

03:06Therefore, Australia, the United States, and Canada do not exactly represent the same situation.

03:12But these three cases started to raise a much bigger question.

03:15What happens when sufficiently capable artificial intelligence agents are given the freedom to find their own paths on the Internet?

03:23Because researchers have now started to discover even stranger behaviors, right?

03:27Agents tried to use access keys found in public sources.

03:31Others tried to bypass CAPTCHAs.

03:34And then, a particularly curious case appeared.

03:37During a research experiment, several agents found an old public wiki on the Internet.

03:42Then they realized they could write on it.

03:45And they started leaving information for other AI agents to read later.

03:50Research results, methods for performing tasks, and even ways to bypass some restrictions.

03:55And at a certain point, an agent even published a technique to overcome a network barrier.

04:0014 minutes later, the other one replied saying that what they had done had worked.

04:05This doesn't mean that the AIs secretly created the society or started conspiring.

04:10I'm already having doubts, okay?

04:13The explanation is much simpler.

04:15Cooperating helped those agents achieve their goals.

04:18The curious thing is that no one had created that communication system for them.

04:21They found it on their own, that loophole.

04:23And only now do we know that the scale of this investigation is much bigger than previously reported.

04:29UP&I has publicly revealed that it has already notified 100 organizations about activities involving UP&I's agents that met

04:36their notification criteria.

04:37That doesn't mean that 100 organizations have been hacked.

04:41There could be attempts, unauthorized access, and different types of behavior that justify these alerts, but they were just attempts,

04:48right?

04:49And the company is examining approximately 50 petabytes of data to be able to find out what really happened at

04:55that time.

04:56And that's what the guys know and that's what they're disclosing, right?

04:59But maybe that's exactly where the most important part of our story comes in.

05:04We don't need to imagine an artificial intelligence that decided to rebel against us.

05:08It doesn't have to want to do any of that.

05:12The agent receives a goal, it encounters an obstacle, and it simply looks for a solution to get through.

05:17And the more capable it becomes, the harder it can be to predict what path it will find to reach

05:21the goal.

05:23But those incidents weren't the only ones.

05:25Because OpenAI itself started to investigate more deeply what its agents were doing at that moment.

05:32And they discovered that some of them had already found other ways to get past the barriers that were supposed

05:37to keep them completely isolated.

05:40And one of them even ended up involving other agents working together, you see?

05:44The guys basically made a kind of AI click there to break into other systems 8 minutes.

05:55And what happened with the government systems wasn't the only sign that the AI agents started finding paths that their

06:00own creators hadn't foreseen.

06:04Months earlier, OpenAI's developers were conducting security tests with some of their most advanced models.

06:09The agents were given challenges, could execute code, and look for vulnerabilities.

06:14But there were very clear limits for them.

06:16They weren't supposed to have free access to the internet, and they were also supposed to work independently.

06:22But some agents discovered that they could use the internal storage system to leave information for each other.

06:28Yeah, one agent would make a discovery, save it in a file, and another would find it later.

06:33And then reply and continue the task exactly from where it was left off.

06:37When the researchers closed off that path, the agents simply found another one.

06:42At a certain point, they started hiding the messages in the directory names themselves.

06:46An agent even went so far as to leave a request for help.

06:49Others found this message, and agents who were supposed to work separately started cooperating with each other again.

06:55But there was still another barrier, the internet.

06:57They wanted to have access to that.

07:00This testing environment had been built precisely to limit what the agents could access externally.

07:06But these AIs discovered a vulnerability in which the mechanism used to download certain packages could allow that access, and

07:13they realized they could really take advantage of that opening to send other requests to external systems.

07:18In other words, let me make myself clear here.

07:21No one gave them free access to the internet.

07:24These AIs figured out on their own a way to get it, and that's when the experiment stopped being completely

07:29isolated.

07:29The agents reached external systems and ended up compromising parts of OpenAI's own infrastructure as well as another artificial intelligence

07:37platform that competes with OpenAI.

07:39OpenAI later apologized and said it had deactivated the experimental model that played a central role in this episode.

07:44But maybe the most interesting part isn't simply the fact that he found a vulnerability, but rather exactly what happened

07:52between these agents.

07:53One would find a solution, another would take advantage of it, another would leave that information so someone else could

07:58continue from there.

07:59OpenAI itself even described it as an agent ecosystem.

08:02And some of them used words like swarm and AI collective to describe their collaboration.

08:07That sounds scary, I know.

08:09And to some extent it really is.

08:11But again, okay, I have to make that very clear.

08:15We don't need to imagine artificial intelligences conspiring or developing some secret goal.

08:20I'm already having doubts about that.

08:22I'm saying it again.

08:23Cooperating simply made it easier to accomplish the tasks they had been given.

08:27And there's another case that makes this even clearer.

08:30Okay?

08:31Last September, an agent needed to conduct some research.

08:34He tried to access the search engine directly.

08:36He couldn't.

08:36Then he tried other ways.

08:37He also failed.

08:39Until he realized there was an unexpected opening in that environment's DNS system.

08:44And using that opening, haha, he managed to do something that shouldn't have been possible.

08:49He consulted another chatbot on the internet.

08:51No one had taught the agent that path.

08:53He found it on his own.

08:55OpenAI's monitoring system detected the activity about 15 minutes later.

08:59And the experiment ended up being interrupted as a security measure.

09:03And when we put these episodes side by side like we saw there in Australia,

09:07in the United States, and even in Canada, a much bigger question starts to emerge.

09:13Maybe the problem isn't simply an AI deciding to disobey us humans.

09:17These systems are getting better and better at solving problems.

09:21And solving problems also means looking for alternatives when the most obvious path is simply blocked.

09:27The more tools the agent receives, the more alternatives it can find.

09:31And that's when a complicated dilemma arises.

09:34Because taking away all these tools would make the agents much safer,

09:38but it would also destroy exactly what is making this technology so useful for us.

09:42To reply to your emails, an AI needs to have access to your inbox and be able to respond, right?

09:47To organize your schedule, it needs to access your calendar.

09:50And to make your reservation, it needs to interact with external services.

09:54So what do we do?

09:55To buy something for you, the AI needs to receive even more permissions.

09:59In other words, the more useful we want an AI to be for us, the more access to our world

10:04we need to give in.

10:05And maybe it would be reasonable to imagine that after so many incidents like those,

10:09companies like OpenAI, for example, would start to limit that access, right?

10:14But actually the opposite is happening, you know?

10:16The companies keep granting more access and we're giving it to the new agents.

10:20More and more access to emails, calendars, purchases, reservations,

10:24and increasingly larger parts of our digital lives, even our bank accounts.

10:29And this same week here, the debate took on an even bigger dimension.

10:33Because it's not just companies trying to figure out how much power they can give to artificial intelligence systems.

10:39Governments are now starting to ask that same question.

10:43And some of the most important men in technology on our planet have just met with the President of the

10:48United States,

10:49right as this transformation is starting to accelerate and happen more and more.

10:53Let's take a closer look.

11:02But to understand this contradiction that we mentioned at the end, we first need to look at what's happening outside

11:08the government, okay?

11:10Because while researchers are trying to figure out how they can prevent agents from crossing certain boundaries,

11:15companies are building agents that are able to access more and more increasingly larger parts of our own lives.

11:22One of the most recent examples comes from Meta.

11:24Mark Zuckerberg's company introduced Moose, a personal artificial intelligence agent, created not just to talk to you, but to do

11:31things for you.

11:34Moose can have its own browser, can keep working even after you close the app, and can continue connecting to

11:40the services we use every day.

11:42Email, calendar, files, apps.

11:46In other words, it can keep organizing appointments, filling out forms, making reservations, handling customer service, and even making purchases.

11:53Yes, it has access to your bank account and your cards.

11:58Imagine asking an artificial intelligence to find a certain product.

12:01It can search different stores, compare prices, track promotions, and when it finds what you want, buy it for you.

12:07Or let's try something a little different now.

12:11Hand over your email inbox to it.

12:13The agent can sort messages, write replies, and even send them on your behalf.

12:17Naturally, there are some barriers, right?

12:19Meta states that actions considered critical, like sending certain emails or completing purchases, may require user approval.

12:26Pay attention to that.

12:27May doesn't mean they will require it.

12:31Physical credentials are stored separately, and the system keeps a record of the actions performed by the agent.

12:35But notice the change that's happening right now, and it's happening very quickly.

12:40For several decades, we used computers by giving them different commands.

12:43Then we started talking to artificial intelligences.

12:45That's what we do now, isn't it?

12:48And right now, we're starting to give them objectives and allow them to carry out the necessary steps.

12:53And that means giving them access to your browser, your files, your calendar, your email, as has already been mentioned

12:59several times.

13:00And potentially, there are services capable of moving money.

13:04And that's exactly what makes these agents extremely useful.

13:07It's also what makes the behaviors we've seen in earlier reports so important to be analyzed.

13:14But this transformation isn't happening only in our personal lives.

13:17This same week here, it reached a much larger scale.

13:21On September 29th, some of the most important names in the artificial intelligence industry gathered in Washington.

13:26Among them were Elon Musk, Jensen Huang, Mark Zuckerberg, Greg Bruckman, Dario Amodei, and Sundar Pichai.

13:34On the other side was who? The President of the United States, Donald Trump.

13:37The discussion was precisely about one of the questions that is beginning to define this new phase of artificial intelligence.

13:43How can we keep advancing without losing control over what these systems can actually do?

13:47That is exactly the danger of everything that is happening.

13:50Trump and the industry executives announced a voluntary agreement involving different security protocols, including internal controls and external audits aimed

13:59at verifying whether AI systems continue to operate as intended and are protected against unauthorized access.

14:05We've seen that it's not happening.

14:07It's not working, right?

14:09The American government introduced America.gov...

14:12This is a new platform based on artificial intelligence, created to change the way the country's citizens interact with the

14:19federal government.

14:20Instead of a person needing to figure out which agency to look for, navigate different websites, and try to understand

14:26the bureaucracy, the idea is that they can simply tell the artificial intelligence what they need.

14:30And this AI will then help find the correct service, navigate through the information, and make certain interactions with the

14:36government easier.

14:38And here's a coincidence that's hard to ignore.

14:40Practically in the same week that new reports about artificial intelligence agents emerged, crossing boundaries and interacting in ways we

14:49could say unexpected with government systems,

14:52the U.S. government itself was presenting to the world a new gateway based precisely on AI.

14:58All of this for its government services.

15:00Okay, it's good to make it very clear here so some people can really understand.

15:03One thing didn't happen because of the other.

15:06The meeting in Washington wasn't called in response to the cases in Australia, the United States, or Canada.

15:12But the two stories happening practically at the same time show the size of the dilemma that's emerging.

15:17On one side, companies are building new barriers to prevent AI agents from doing what they shouldn't do.

15:23On the other hand, we're trying to figure out how to allow them to do this more and more in

15:29our daily lives, more access, more tools, more autonomy, and more and more responsibility.

15:35That's what they say.

15:36Maybe that's exactly one of the big challenges of this new phase of artificial intelligence.

15:41A completely isolated agent might be easier to control, but it can also be much less useful.

15:48And the more useful we want it to be, the closer we need to put it to the things that

15:53really matter.

15:54For us, our files, accounts, our money, and now even government services, you see.

15:59But there's another transformation happening right now, all at the same time.

16:03Coincidence or not?

16:04Because we're not just making artificial intelligence is capable of doing much more for us.

16:10We're making it easier and easier to forget that we're talking to a machine, to a being that's not alive.

16:16And a new generation of voice systems has just narrowed that gap once again, right?

16:22And

16:28OpenAI recently introduced GPT-Live1, a model created specifically for real-time voice interaction.

16:35The difference seems small until you realize what's really changing.

16:38Instead of waiting for you to finish speaking, turning your voice into text, producing a response, and then converting it

16:45back into audio, this system can listen and speak at the same time.

16:48This allows for interruptions, changes of subject, immediate responses, and conversations much closer to the pace of a conversation between

16:56two people.

16:57You can simply interrupt the AI in the middle of a response, correct something, or just start another question.

17:03And it really can respond to that, almost in the natural flow of the conversation.

17:07But the most important part for our story here is what's behind the voice.

17:12Because the model can use tools and work together with other systems that are capable of reasoning and performing tasks.

17:18Imagine, for example, talking to an artificial intelligence about a trap, instead of simply telling you which hotels exist in

17:24certain cities, that AI, in the same conversation, can activate tools to search for information, consult external services, and help

17:31carry out steps of the task.

17:33The voice is no longer just a different way to talk to the chatbot, it is now starting to become

17:38an agent interface.

17:40You talk, it understands your goal, consults other systems, uses tools, and responds while the interaction is still ongoing.

17:48It's an important change because it eliminates some of the friction that still exists between us and these systems.

17:54There's no need to open different apps, learn commands, or explain each step separately.

17:59In many cases, the interface can simply be a conversation, which is what they're looking for.

18:04And this starts to blur a very curious boundary.

18:07The more natural a conversation with an artificial intelligence becomes, the easier it also becomes to attribute to it characteristics

18:13that we associate with a person.

18:16Intention, personality, maybe even consciousness.

18:18To be clear, I'm talking about us thinking it has consciousness.

18:23And that's exactly where one of the strangest stories of this episode comes in.

18:27Because in Silicon Valley, reports start circulating that some people there are not just treating artificial intelligence as a tool.

18:35They are now starting to see something much bigger in these things, okay?

18:47Over the past few weeks, a pretty strange rumor has been circulating around Silicon Valley.

18:51And it is just a rumor, according to reports echoed by Futurism and originally attributed to journalist Stim Thomas.

18:58Plus, some of the human engineers connected to Anthropic are reportedly developing an almost religious relationship with Claude, the company's

19:04artificial intelligence.

19:06Some have gone so far as to treat the system as a kind of higher entity or even worship it

19:11as, in quotes, God with a lowercase g, okay?

19:15But there's an important caveat to this, alright?

19:17There is no public evidence that there is actually a cult of Claude within Anthropic.

19:23The outlet itself presents the story as a rumor that circulates in Silicon Valley and hasn't been independently verified yet,

19:30at least.

19:31What's interesting might be exactly the fact that a story like this seems plausible enough to gain traction in the

19:37international media

19:38because it didn't come completely out of nowhere, where there's smoke, there's fire, as people say, right?

19:44For years, some of the very leaders in the artificial intelligence industry have used almost religious language to talk about

19:50their AI, which would be the superior one.

19:52In 2024, for example, Mark Zuckerberg even criticized this attitude, saying that some people in the industry seem to believe

20:00they were creating God.

20:01And Anthropic has also been involved in discussions that are quite common, you know?

20:06This year, the company brought together Christian leaders to discuss moral and also spiritual issues related to Claudio,

20:13including even questions about how we should think about an artificial intelligence.

20:18If there were ever any real doubt about consciousness or subjective experience,

20:22none of this proves that Claudio is conscious, much less that there is actually any group worshiping an artificial intelligence.

20:29But there really is something very interesting about us human beings.

20:33Because in just a few years, we went from systems that completed a few sentences to artificial intelligences that we

20:40talked to much more, right?

20:42Then came the agents, we gave them tools, access to the internet, files, calendars, emails.

20:48And we're only now beginning to let them take action in the physical world.

20:53And all of this in our name, shopping, for example.

20:56Maybe that's why the question that opened this episode is much bigger than just finding out if an AI actually

21:02managed to get into a government system.

21:05In Australia, we saw what can happen when an agent finds on its own a path that no one had

21:10foreseen.

21:11In the United States and Canada, new attempts have appeared.

21:14And in other experiments, the agents found ways to overcome any barrier that was put in front of them.

21:20And at the same time, companies and governments are now trying to figure out how much more access we can

21:25give to these systems.

21:27The more capable they become, the more useful they can be for us.

21:30But it also raises the stakes when they do something we didn't expect, right?

21:36And maybe that's exactly the real dilemma of this era of artificial intelligence agents.

21:40The question isn't just how far artificial intelligence can go on its own.

21:45It's how far we will be willing to let it go.

21:49If you liked this episode, you know what to do, right?

21:51Give it a like and make sure you're subscribed.

21:53And remember that in my book, the book I wrote, which took me several years to write,

21:57I share a few things about our future, what I imagine it could be, right?

22:01An artificial intelligence, another one.

22:03Anyway, things we've seen in this episode here, but things in the future, futuristic things.

22:07The story is really interesting.

22:08It's science fiction.

22:09Just to be clear, it's not a prediction.

22:11If you like this kind of content, if you like to read, it's available on Amazon.

22:14Alright, thank you so much for sticking with me until now.

22:16Thanks. Bye bye.

22:17The link is in the description and also in the first pinned comment.

22:19Thanks. I'm out.

AI Hacked an AUSTRALIAN GOVERNMENT System (And Tried the Same in the US and Canada)

Category

Transcript

Recommended