OpenAI made first known AI hack of a govt system

A new incident involving. agents going rogue. Australia's prime. minister says an open. AI agent. hacked into a database of the country's. national health care. system, accessing public and nonpublic files. The incident happened in June, but OpenAI says it is. only became aware of it in August. The prime. minister said he spoke to OpenAI. CEO at the UN and expressed Austria's. Australia's, quote, extreme concern. CNN's Hannah Montgomery live in Tokyo.

with more on this. It sounds pretty concerning, Hanako. Yeah. Alex, as you just said, the Australian prime Minister, Anthony Albanese, said he expressed. extreme concern. to Sam Altman, the CEO of OpenAI, directly. And as you stated also, Alex, this is the first known case. of AI hacking a government network. But it's also the latest. example of AI agents acting in ways. that their human operators didn't intend. This is the Australian prime minister,

Albanese, explaining to reporters. what exactly happened. during a sideline. meeting at the United Nations. General Assembly on Wednesday. Listen to this. No personal information. is believed to have been accessed. at this stage, but investigations are ongoing. Evidence currently. available is. there is no broader. compromise to the Services. Australia network. Nonetheless, the situation is obviously unacceptable.

Now, as you heard Albanese. say, there's. currently no evidence. that any personal information. was breached. But we do know that the AI agent got into. Australia Medicare statistics database. We also know that IT access. to both public. and nonpublic files, including aggregate health. statistics and internal file names. And according to an OpenAI spokesperson, the agent. even wrote files into the system. Now, OpenAI said that the security breach. actually happened back in June,

when one of its agents was conducting. research into health care. expenses and spending. when one of its agents actually bypassed. some security blocks. and got access to information. that it wasn't authorized to do so. But the company only learned of this. security breach two months later. Back in August, when it was conducting. checks on its models. Now, the Australian government. was actually only informed, though, of this breach on September 10th, when a notice was sent. to a generic inbox.

So that's. basically three months. after the security breach. actually happened. Now, Albanese. expressed concern over. this delay as well, saying that. essentially it's not acceptable. And given the severity of the situation, the government needed to be informed. much sooner. Now, looking ahead, we do know that the Australian government. is launching a forensic investigation. They are also considering a lawsuit, and Albanese. has said that. potentially three other government. networks. were also affected. by this security breach.

And again, Alex, just for some context, I mean, we again, know. that this is the first case. of a government network. being hacked by AI. But it's certainly not the first time. OpenAI's agents have hacked a company. We know back, earlier this year when essentially, AI agents hacked an AI company, hugging face. We also know again. that the topic of. including more guardrails for. AI is certainly being discussed. at the United Nations General. Assembly this week, as more and more.

AI CEOs. and governments are urging. for stronger guardrails and, again, just more protocols. in terms of using AI more safely. Alex. Yeah, the first government. hack that we know of, and this even happened months ago. So we don't know if more have happened. even since then. Joining me right. now is a chief global affairs. officer for OpenAI. Chris Lehane Chris, thank you for coming in. The Prime Minister of Australia says that.

how long it took you, the OpenAI, you all to notify. them is unacceptable. It happened June 18th. They were notified September 10th. Do you think that gap between incident. and reporting is an acceptable gap? Yeah. Let me just give the context. So I think as your reporting indicated, there was an incident. that took place in. July that's known. as the hugging face incident. After that, we became aware of that, we OpenAI, initiated an ongoing review.

of that review continues. where we've been going backwards. and trying to see, you know, what else we could learn. As we've gone through that process. and as we've come across other incidents. And to be clear, the other incidents are not at the scale. and caliber. and kind of what the hugging face was. We have then taken that information. and then communicated. that and conveyed it. to the impacted parties. And we did this several weeks ago. to the Australians. We've we do leave it up.

to the impacted entities to decide. whether they want. to make this public or not. And we do seek to work with them. to be able to identify. and the exploitation. of the vulnerabilities in their code. that would have made them. exposed to this. And, you know, this is an ongoing review. We continue to communicate things. as we come across them. We take it very, very, very seriously. The other point. I would just really emphasize in. all of. this is after the initial event, the Hugging Face event in July, you know, we did take a step back. We actually put a pause. on some of our training,

and we endeavored to build. in the types of safeguards, both on something. called alignment, making sure the technology's. actually following the directives. of humans. and the guardrails, making sure. it's safe when it is deployed. And since then, we've been able. to release other models. We actually have a model out right now, which is called Astra, which is both. the most capable model in the world, but also a model. that has been put out in a way. that's aligned and safe, for public use. So you telling Australia, you guys finding this in August.

and telling Australians. on September 10th, is that gap between finding it. and reporting it, especially to a government? Is that an acceptable. gap or are you shortening that? Yeah, I think the gap, you know, as we've discovered, this stuff is again, an ongoing review. when we see it, we then analyze it. and then we communicate that to. the Australians. And again, I think we communicated this. to the Australians, several weeks ago, obviously have incredible respect. for the Prime Minister.

He obviously expressed himself yesterday. But I do think that the protocol. that we are following, which is we have an ongoing review, we identify this. Once we identify it, we then convey that and communicate it, to the impacted party. In this case, I think that. if folks have known about it. for several weeks, so, so not being told about it, what what. the Prime Minister says, when you guys inform them. that time period, you're disputing that. No, no, no, I want to be, as I said, exceptionally respectful. of the Australian government.

and of of the Prime Minister. I'm just sharing with you. what our protocol is. and our process and our processes. We do have an ongoing review. When we find the information, we convey. that information. in this case. to a number of parties in Australia. And then we leave it. up to them in terms of who. they want to communicate it to. and how they want to share it, including. whether they want to share it publicly. I think another important aspect. in all of this is, you know, these are entities. that are involved in health care issues. And so I think amongst the questions. that we have gotten was, you know, where any individuals.

are people impacted by this. and there is no personal. health care information. that was, impacted. by this particular incident. But again, this is really serious, right? We understand that is serious. That's why we have an ongoing review. The review is sort of working backwards. And as we identify information, we do convey it. and communicate it. to the impacted parties at the UN. Sam Altman talked about. if AI is to be democratic, the most important decisions cannot be. made by labs in San Francisco alone. You're calling for a global body to set.

international standards. when it comes to AI. President Trump. has rejected the creation. of that global body, even saying, today, I want to leave it exactly where it is. That's China's position. Also, our guardrail is the justice Department. If you cannot get buy in. from the United States government. and the Chinese government, do any global standards have any teeth? Yeah. So our. perspective and Sam shared. this obviously yesterday at the UN. is that, you know, this technology.

is really a first technology. in the modern era. of this kind. that has really been developed, deployed, and released by the private sector. There's incredible upsides. Sam talked yesterday about there's. an opportunity here. for this technology to really empower. individuals, potentially spark, you know, a modern day. new, new renaissance. where really individuals can participate, through this technology. in the economy and in other ways. But it also comes with, you know, incredibly important issues.

that we need to address. on risk and on safety. And the US right now. is the leader in this technology. We are the innovation leaders. We do believe it's incredibly important. that there's. this technology. moves forward, particularly. if we want to get that Renaissance impact. that you do. build it out on the democratic standards. around the world. And so there's a real opportunity here. for the US to use the lead. that it currently has, right. An opportunity for the president. to present. to continue to be the innovation leader. Yeah, right. But Chris, if the president doesn't want to.

doesn't want to set. standards, doesn't. want to take part in a global body, do any of those standards. that any the rest. even of the rest of the world. without China. and the United States would set. Does it have any teeth? Yeah. I think there's, an opportunity here. to take some steps forward. in that process. that one of the things. that we have talked about is. could there be some equivalent. of a red phone, an actual just basic mechanism. so that the two countries. can actually be in contact. with each other on this technology?

It doesn't necessarily mean. we're going to get standards. today or tomorrow or next week, but the ability to start to create. some type. of constructive dialog. and interaction between the two parties, I do think would be a step. in that direction. People have generally talked. about that concept. I'm old enough to have remembered, during the middle of the Cold War, when things were much, much more intense, between the US and the Soviet Union. than anything that exists today. There was a red phone, and both countries. were able to stay in contact. with one another.

And so, you know, perhaps that's the first step. that we can take here. Ultimately, I do believe that, you know, this is incredibly powerful technology, and it would benefit the world. and certainly benefit U.S. leadership to ultimately have. those international standards in place, particularly led by the US. I think a lot of people. would agree with you, Chris. One thing on the business side, there's a saying that. when China walks in the door, profits leave. It's very clear. that China is very much in the room. when it comes to large language models. With that in mind, I mean, how do you ever turn a profit?

Well, you know, for open AI, right? We today over a billion people. who are using this, on a regular basis, you know, we have a new model. that we recently released. that I referenced earlier, Astra, which is doing fantastic. And ultimately, you know, for companies like ours. where we have a diverse set of revenue. sources, you know, but we need to do is to continue. to develop and innovate. and put products out there. that are going to actually help people. in their daily lives,

make their lives better. And we are seeing that with the products. that we are releasing. and the growth that we've experienced. as a result of that. Chris Lehane of Open AI, it's great to have you come in. Thank you so much for your time. Joining me right now is CNN tech. editor Lisa. at a check up for more on this. What did you hear here? What do you want to give me your take. on what we heard from Chris Lehane. and OpenAI and all this, especially when it. first when it comes to. the Australia incident. Yeah. Kate, the one thing. that really jumped out to me. is this mention of an ongoing review. as part of this process.

that Lehane talked about. And I think. one of the big questions around. regulation is kind of the transparency. behind it. You know, what exactly does. an ongoing review entail? Now, OpenAI has been fairly public. about its processes. It does publish a lot of blog posts. kind of outlining this, but I think that's kind. of the key question. with when it comes to regulation here. is that right now, this is up to OpenAI's discretion, how it reviews these incidents, how when it decides. to disclose that to third parties. And clearly. that is a key, cause.

for concern and a care topic. A key. topic here is how long it took OpenAI. to disclose this. So I think that's part. of this real push here. to kind of come up. with some kind of global standard, some kind of industry wide standard, some form of regulation. So that maybe it's. not just up to these companies. that are developing the technology. to decide. when they disclose things like that. And OpenAI. did say recently. that they're going to disclose. incidents more frequently, but that seems to be more focused. on incidents that it discovers.

during testing and stuff like that. So I guess we'll have to kind of watch. for that moving forward. How often they come out and disclose. these incidents? Yeah. What's the definition of will disclose. how long. What is the definition. of transparency even because Australia is. very obviously not happy. with the level of transparency. or the, you know, the timing. that they've got. in terms of that incident. have to remember globally, like the perspective on technology. and consumer privacy. And all of these things are so different. globally. Europe has a much different perspective. than the US, for example. So that's another challenge as well.

Add it to the list. It's good to see Elisa. Thank you.