Television Interview, News24

SUBJECT/S: Artificial intelligence

HOST, PETER STEFANOVIC: Well, let's try and get some more answers on what's happened with this breach this morning. As you know, folks, we've been trying to reconnect with the Prime Minister, but his shot keeps freezing and it's not coming through. Anyway, we've got the Acting Prime Minister, Richard Marles with us now to try and work out what's going on. Richard Marles, thank you so much for your time. So, can you just elaborate with what we're dealing with here. What has happened?

ACTING PRIME MINISTER, RICHARD MARLES: Well, we've been notified by OpenAI that its agents within its artificial intelligence model gained unauthorised access to a medical statistics portal of Services Australia. This occurred back in June. Obviously, this is unacceptable and we've expressed our concern to Sam Altman. The Prime Minister has done that overnight. Sam Altman, the CEO of OpenAI, we've made clear that this is an unacceptable situation. We've established a task force which will be led by PM&C to look into this in terms of what's happened in relation to our systems. We are working with OpenAI in relation to that and that has been- we have been working cooperatively with them to understand exactly what has occurred here. I think it's important to stress that there's no data of any person which has been accessed here. As I'm advised, there's no impact to the system. So, actually, as I am advised, the impact to our systems is very minor. That said, what we're seeing here is an unauthorised access that's been occurred via an AI model. And that is completely unacceptable. And it's really, I think, something of a warning as to how we need to be globally developing AI very carefully. It's why the Prime Minister in New York, along with other world leaders, has put out a call to make sure that AI developed in a responsible way. We've seen the tech company CEOs address the United Nations overnight, making clear that this need, this is a technology which does offer so many benefits to humanity, but must be developed with extreme care. And we have been at the forefront of ensuring that.

STEFANOVIC: We will get to all of that but just to back things right up, I just want to try and get some specifics. How would you define an AI model? What does that mean for ordinary people?

MARLES: Well, I'll do my best here and try to answer that. OpenAI has a model which will have agents which are part of the model which will go out and find information. As I understand it, in this instance the task that was given to the model, to the agents was benign. It was around trying to do a research task in medical and health statistics. In the course of that it's gone across the internet if you like, to try and find out that information. It has come to the medical statistics portal of Services Australia and in seeking to get that information has asked questions, information has not been given up by the portal and then it has worked its way around and gained an unauthorised access to that portal. And that's really what is of concern to us here and that's what we've expressed that concern to us.

STEFANOVIC: But that's, but that's non-human. Non-human.

MARLES: Oh yes, to be completely clear, that is non-human. Sorry, I should made that really clear. This is an AI model. This is, this is a- these are non-human artificial intelligence agents which are doing this. They have been set, as I say, a relatively benign research task to research medical and health statistics. But the agent, a non-human agent, has come to the medical portal of Services Australia, asked questions, been told that's not information that's going to be provided here. And then has worked, actually worked its way into the system in an unauthorised way and gained information in any event and that is what concerns us.

STEFANOVIC: I mean that in itself is quite frightening within its own capability, isn't it? Just to do that, to operate on its own, to think on its own around side our own parameters.

MARLES: Yeah, I mean this is obviously the capability that AI represents and we've seen other incidents of this occur elsewhere but this is the first time we've seen this in respect of an Australian government website, which is why we are particularly concerned here now. I mean, this is the Australian government's own information owned website which has been had an unauthorised access to it but been done by an AI agent. But this is the capability which is out there and it's obviously why we have expressed our concern with it and why, whilst the impact is minor in the sense that no person's data was was accessed here, we're not talking about a system which has been affected in any way. It is still a very serious thing that we see an unauthorised access into an Australian government site and as a result of that we're taking this very seriously.

STEFANOVIC: And so you found out this happened in June- Am I right, June?

MARLES: Yes it did.

STEFANOVIC: And when did you find out about it?

MARLES: So, OpenAI themselves found out about it in August, as they were going through and looking at all that its model had done, that the agents had done. Services Australia were informed about this around two weeks ago. Government Ministers were informed about this at the end of last week and over the weekend. That's when I became aware of it. Since then, what we've been trying to do is to really make sure that we have all the facts at hand to, you know, be clear about the impact itself-

STEFANOVIC: Why the delay? That's three months, Minister. Why the delay? Why are you announcing that now?

MARLES: Well, in terms of OpenAI engaging with the government, that is a point that the Prime Minister has made to OpenAI in terms of wanting why there was a delay in being informed by OpenAI. As we have been advised by them, they were not aware of this until August. At our end, you know, over the last few days, what we've sought to do is to be completely clear as to exactly what's happened, to make sure that the initial advice that the actual impact relation to the site and data was relatively minor. It was important that as we announced this publicly, we were able to make that clear to give some sense of reassurance about the actual impact of this incident. And that's why we've, we've spent the last few days assuring ourselves of that before making this public. But the question you raise in respect of why it took so long for the Australian government to be notified, is a reasonable one and that's a point that we've made to OpenAI.

STEFANOVIC: Well, cynics might argue that this has been timed, this announcement today is timed with the Prime Minister's appearance at the UN today. Is there anything in that?

MARLES: No, there's nothing in that. I mean, as I say, we became really aware of this at a Ministerial level over the course of the weekend and we've been spending the last few days to make sure that we can, in making this public to the Australian people, be clear that the impact of the incident is minor. But that the significance of it in terms of an unauthorised access to an Australian government website, that's a serious issue. And we want to- there is still more work to do-

STEFANOVIC: Is that a national security problem then that it’s been able to get around our system?

MARLES: I mean, so I think that's a reasonable question. And again, these are the issues that we are wanting to look into. You know, we have a number of systems obviously across government. They have varying levels of security associated with them. Our, you know- 

STEFANOVIC: Yeah, but this has broken through. 

MARLES: Yes, but to be clear, Pete we're talking about a medical statistics portal which would not have the same level of security associated with it that we operate in relation to our most sensitive information. But we are looking across the board in respect of the security that we have in place and that's why we've established the task force to look at this in all its detail. I emphasise again that the information that has been sought here is medical statistics. It's not individual’s data. So it's relatively low level in terms of its sensitivity. And the security around that site wouldn't be anything like what we have in relation to our sensitive data. But there's no question here that we've seen an unauthorised access into an Australian government website that is a serious concern. As you say, it's not, this is not a human. This is an AI agent which has done this. And that's why we're taking this very seriously. 

STEFANOVIC: Could this possibly be a problem for OpenAI investing here in Australia, which is what it wants to do with data centres and the like?

MARLES: Oh, look, I think we take everything a step at a time here. I mean, I think there are a range of issues around how AI develops. Right now what we're doing is being focused on this incident and what this says around the development of AI models and how they are being and how they're operating. And I think when we're thinking about this incident, think about it in the context of the essay that Dario Amodei wrote a couple of weeks ago. But most significantly, the statement that the Prime Minister, along with other leaders put out in New York this week, making sure that AI is developed in a responsible way with guardrails, where, you know, there is the safety measures around AI, well ahead of the capability that is being built. So, that this is a technology which is developed for humanity and not one which puts humanity at risk.

STEFANOVIC: So, Donald Trump, the US President, he doesn't have a problem with all of this. He doesn't think there needs to be regulations. So, there appears to be this divergence of views between the US President and the tech titans themselves. So, who's right here? Is the President right? Are the tech bosses right?

MARLES: Well, look, I think that what the tech CEOs are saying is important to listen to. The essay that was written by Dario Amodei, I think was a very significant moment but so too is the appearance of the tech CEOs before the United Nations overnight, where again, that same message was reiterated. But from an Australian perspective, our position is really clear because the Prime Minister was, was leading other countries in New York this week in making a statement that AI needs to be developed with extreme care. It needs to be done in a way where the safety around it, the guardrails that are put in place are well in advance of the capability which is being developed. And that is precisely so that this is a capability that can be working for humanity and not one which is engaging in unintended actions. And whilst in a sense what we've seen happen in Australia has, you know, a very minor impact, it is obviously concerning when you think about what has occurred here and it's something of a warning.

STEFANOVIC: Well, I'll close on this point and given that you said this is a minor incident but does have ramifications but big picture, I mean you've got some people who think that unchecked AI could lead to the end of humanity. Now there have been a number of episodes in the past, doomsday scenarios that just haven't happened. So, what do you say to those who believe well this might just be another episode of fear mongering by governments and elites?

MARLES: Well I think we need to be careful in the language that we use. But, we also-

STEFANOVIC: But is the end of humanity overplayed here?

MARLES: Again, think we need to be very careful about the language that we use and that's not language I'm using. Look, this is a very powerful capability, a very powerful technology. It needs to be developed with extreme care. To be clear, we are treating this incident very seriously. That an artificial intelligence agent, a non-human, has gained unauthorised access into a government website is something which is extremely concerning and whilst the impact of that might be minor in this instance, it is a warning in terms of what can happen going forward. And that's why this needs to be developed with extreme care.

STEFANOVIC: Now lengthy chat I know you've got to go to another interview but the Acting Prime Minister Richard Marles really appreciate your time as always this morning to cover all those breaking details.