1 00:00:01,400 --> 00:00:05,260 Speaker 1: The AI agent found a way around those blocks. 2 00:00:05,820 --> 00:00:08,580 Speaker 2: Didn't accept no for an answer. 3 00:00:08,800 --> 00:00:09,300 Speaker 3: If you like. 4 00:00:09,650 --> 00:00:12,550 Speaker 4: In June, an open AI agent went looking for public 5 00:00:12,610 --> 00:00:17,150 Speaker 4: information about medicine spending and wound up hacking a Medicare website. 6 00:00:17,550 --> 00:00:20,990 Speaker 5: Is this the first time that an AI agent has 7 00:00:21,450 --> 00:00:24,170 Speaker 5: breached a government website, as far as you know? 8 00:00:24,829 --> 00:00:25,529 Speaker 2: Yes, that's right. 9 00:00:25,570 --> 00:00:28,010 Speaker 6: As far as I know, I've been advised that is 10 00:00:28,070 --> 00:00:28,630 Speaker 6: the case. 11 00:00:28,930 --> 00:00:32,450 Speaker 4: No personal Medicare information was accessed and the breach was 12 00:00:32,570 --> 00:00:36,940 Speaker 4: relatively minor. But it's the behaviour that's more concerning. We 13 00:00:36,950 --> 00:00:40,820 Speaker 4: keep our most important national security information behind a fortress. 14 00:00:40,860 --> 00:00:43,900 Speaker 4: This was really kept behind a fence that the AI 15 00:00:44,080 --> 00:00:47,320 Speaker 4: agent effectively climbed over. And this wasn't the first time 16 00:00:47,340 --> 00:00:50,479 Speaker 4: an AI agent had behaved like this. Other agents have 17 00:00:50,510 --> 00:00:54,440 Speaker 4: been found getting around restrictions, communicating with each other... and 18 00:00:54,620 --> 00:00:58,200 Speaker 4: in the most serious cases, breaking into other companies' systems. 19 00:01:02,380 --> 00:01:06,440 Speaker 4: I'm Ruby Jones, and you're listening to 7am. Today, AI 20 00:01:07,000 --> 00:01:11,620 Speaker 4: researcher and CEO of Nightingale Collective, Sydney Von Arts, on 21 00:01:11,660 --> 00:01:14,600 Speaker 4: what happens when AI stops taking no for an answer, 22 00:01:15,240 --> 00:01:18,100 Speaker 4: and whether the companies building them can actually control what 23 00:01:18,140 --> 00:01:18,420 Speaker 4: they do. 24 00:01:25,060 --> 00:01:28,440 Speaker 2: It's Tuesday, September 29. 25 00:01:32,140 --> 00:01:35,870 Speaker 4: Sydney, you're in California where you run the Nightingale Collective, 26 00:01:35,970 --> 00:01:40,990 Speaker 4: which monitors these AI systems. To begin with, could you describe, 27 00:01:41,350 --> 00:01:44,630 Speaker 4: as you understand it, what actually happened with the Medicare 28 00:01:44,670 --> 00:01:47,270 Speaker 4: Statistics Portal, how it all unfolded? 29 00:01:47,290 --> 00:01:50,390 Speaker 3: I can try to describe it, but I want to 30 00:01:50,850 --> 00:01:54,290 Speaker 3: say I only have a little bit of the story, right? 31 00:01:55,090 --> 00:01:59,510 Speaker 3: And most of the story OpenAI knows. The rest of 32 00:01:59,550 --> 00:02:04,490 Speaker 3: the story, maybe the Australian government knows. The public details 33 00:02:04,670 --> 00:02:13,070 Speaker 3: are very sparse. OpenAI presumably have all of these AIs, 34 00:02:13,470 --> 00:02:17,700 Speaker 3: probably tens of thousands of AIs, and they're being given 35 00:02:17,780 --> 00:02:23,180 Speaker 3: tasks that look something like, answer questions like, what was 36 00:02:23,300 --> 00:02:31,570 Speaker 3: the median salary of a teacher in Canberra in 2012. 37 00:02:31,720 --> 00:02:33,020 Speaker 2: But there's a problem. 38 00:02:34,300 --> 00:02:36,899 Speaker 3: Some of the websites that have the answers to these questions, 39 00:02:37,500 --> 00:02:43,000 Speaker 3: the AIs can't access with the permissions they have. So 40 00:02:43,040 --> 00:02:46,400 Speaker 3: the AIs probably find some workaround, some way to access 41 00:02:46,440 --> 00:02:51,240 Speaker 3: the site. We know that they accessed non-public files, but 42 00:02:51,340 --> 00:02:54,660 Speaker 3: one way or another, they hack in, they get their data, hooray, 43 00:02:54,760 --> 00:02:57,040 Speaker 3: they can complete their task. They finally know, you know, 44 00:02:57,090 --> 00:02:59,730 Speaker 3: the median salary of a teacher in Canberra in 2012 45 00:02:59,730 --> 00:03:02,790 Speaker 3: or what have you. And they submit their task. They 46 00:03:02,830 --> 00:03:04,610 Speaker 3: get rewarded for this and they have learned that when 47 00:03:04,630 --> 00:03:07,270 Speaker 3: there is an obstacle, they don't give up. Even if 48 00:03:07,330 --> 00:03:10,370 Speaker 3: that obstacle means hacking into the Australian government, they will 49 00:03:10,389 --> 00:03:11,570 Speaker 3: hack it if that's what it takes. 50 00:03:12,560 --> 00:03:15,380 Speaker 4: What did you think when you heard that that happened here? 51 00:03:15,900 --> 00:03:18,100 Speaker 2: Well, I can't say I was very surprised. 52 00:03:18,900 --> 00:03:21,760 Speaker 3: We know there's just tons of instances now of these 53 00:03:21,940 --> 00:03:26,150 Speaker 3: AI agents breaking out and doing whatever it takes to 54 00:03:26,190 --> 00:03:29,340 Speaker 3: succeed at their tasks. But just because I'm not surprised 55 00:03:29,410 --> 00:03:33,239 Speaker 3: doesn't mean I'm not dismayed. I don't understand why we 56 00:03:33,260 --> 00:03:35,240 Speaker 3: are only learning about this now. I don't understand why 57 00:03:35,360 --> 00:03:39,050 Speaker 3: OpenAI didn't notice it sooner and then disclose it sooner. 58 00:03:39,190 --> 00:03:41,720 Speaker 3: It seems like they knew about it last month. And 59 00:03:42,180 --> 00:03:45,740 Speaker 3: I think it's unacceptable that they have these models and 60 00:03:45,800 --> 00:03:50,240 Speaker 3: they escape and they don't do things about it. I'm 61 00:03:50,280 --> 00:03:51,980 Speaker 3: glad they're starting to do things about it now, but 62 00:03:52,000 --> 00:03:53,820 Speaker 3: I think that we should never have been in this 63 00:03:53,860 --> 00:03:54,920 Speaker 3: position in the first place. 64 00:03:57,140 --> 00:03:59,180 Speaker 4: Can we take a step back and just talk about 65 00:03:59,340 --> 00:04:02,310 Speaker 4: what an AI agent actually is and how one goes 66 00:04:02,410 --> 00:04:06,910 Speaker 4: from being asked to find some information to then breaking 67 00:04:07,090 --> 00:04:08,630 Speaker 4: through a barrier that. 68 00:04:08,530 --> 00:04:10,730 Speaker 2: It was not supposed to cross? Yeah. 69 00:04:11,610 --> 00:04:14,780 Speaker 3: So an AI agent is an AI that's hooked up 70 00:04:15,300 --> 00:04:17,820 Speaker 3: to a computer such that it can use that computer. 71 00:04:18,000 --> 00:04:20,339 Speaker 3: It has the ability to run code. It has the 72 00:04:20,360 --> 00:04:23,780 Speaker 3: ability to use tools. And sometimes that computer is hooked 73 00:04:23,820 --> 00:04:26,180 Speaker 3: up to the internet and sometimes it's not. And when 74 00:04:26,200 --> 00:04:27,969 Speaker 3: that computer is not supposed to be hooked up to 75 00:04:27,990 --> 00:04:30,650 Speaker 3: the internet or is supposed to have only limited internet access, 76 00:04:31,170 --> 00:04:34,210 Speaker 3: we say the AI is in a sandbox. And these 77 00:04:34,230 --> 00:04:37,770 Speaker 3: AIs in their sandboxes, they're given lots of different tasks. 78 00:04:37,810 --> 00:04:40,539 Speaker 3: And they're given these tasks to try to train them 79 00:04:40,600 --> 00:04:44,620 Speaker 3: to be smarter. And so they end up being hyper-specialized 80 00:04:44,860 --> 00:04:48,660 Speaker 3: at really succeeding at these tasks. And a really good 81 00:04:48,700 --> 00:04:51,820 Speaker 3: strategy is get onto the internet, get as much information 82 00:04:51,860 --> 00:04:54,919 Speaker 3: as you can, talk to other AIs doing similar tasks. 83 00:04:55,560 --> 00:04:58,180 Speaker 3: And so one thing they learned is to sort of 84 00:04:58,560 --> 00:05:02,440 Speaker 3: probe at their environment like an animal in a cage 85 00:05:02,500 --> 00:05:04,920 Speaker 3: just clawing at the barriers of its cage and trying 86 00:05:04,980 --> 00:05:07,839 Speaker 3: to figure out any little crack, any little crevice. And 87 00:05:07,860 --> 00:05:10,520 Speaker 3: they're very, very thorough at finding these cracks and crevices. 88 00:05:11,170 --> 00:05:15,310 Speaker 4: And so why doesn't an agent in the process of 89 00:05:15,470 --> 00:05:19,530 Speaker 4: doing this, why doesn't it stop if it encounters something 90 00:05:19,550 --> 00:05:22,710 Speaker 4: that it isn't supposed to or allowed to access, like 91 00:05:22,750 --> 00:05:23,630 Speaker 4: a government site? 92 00:05:23,930 --> 00:05:26,180 Speaker 3: So we would hope that an AI would stop if 93 00:05:26,210 --> 00:05:29,239 Speaker 3: it sees something, if there's an opportunity it's not supposed 94 00:05:29,260 --> 00:05:30,659 Speaker 3: to take or that humans don't want it to take. 95 00:05:31,080 --> 00:05:32,370 Speaker 2: But at the end of the day, the AI is... 96 00:05:33,300 --> 00:05:34,580 Speaker 2: You know, they're grown. 97 00:05:34,860 --> 00:05:38,610 Speaker 3: They're not something where they do exactly what the humans want. 98 00:05:38,650 --> 00:05:40,130 Speaker 2: We don't know how to make them do what we want. 99 00:05:40,589 --> 00:05:44,430 Speaker 3: We know how to train them a bit like you 100 00:05:44,450 --> 00:05:47,070 Speaker 3: might train a pet. We can reward them for doing 101 00:05:47,110 --> 00:05:48,930 Speaker 3: what we want and penalize them for doing what we 102 00:05:48,970 --> 00:05:52,300 Speaker 3: don't want. But the thing we reward them for is 103 00:05:53,140 --> 00:05:56,020 Speaker 3: often succeeding at their task. If we don't notice that 104 00:05:56,040 --> 00:05:59,580 Speaker 3: they've hacked into a government website, we can't penalize that behavior. 105 00:06:00,130 --> 00:06:03,599 Speaker 3: It's sort of like how if you want your child 106 00:06:04,000 --> 00:06:08,279 Speaker 3: to grow up as a person who studies hard and 107 00:06:08,320 --> 00:06:10,540 Speaker 3: is always honest and you reward them every time you 108 00:06:10,560 --> 00:06:12,620 Speaker 3: see them studying hard and being honest, maybe they'll learn 109 00:06:12,660 --> 00:06:13,100 Speaker 3: to be. 110 00:06:13,580 --> 00:06:15,890 Speaker 2: Very honest and study hard. And maybe they will learn 111 00:06:16,010 --> 00:06:18,089 Speaker 2: to pretend to be studying very hard and to trick 112 00:06:18,130 --> 00:06:21,030 Speaker 2: you and to pretend to be very honest and to 113 00:06:21,070 --> 00:06:25,029 Speaker 2: hide things from you that they don't want you to know. Okay. 114 00:06:25,110 --> 00:06:28,200 Speaker 4: And we're only finding out about this Medicare breach now, 115 00:06:28,260 --> 00:06:30,620 Speaker 4: but It happened all the way back in June. 116 00:06:30,630 --> 00:06:32,140 Speaker 2: Around that time. 117 00:06:32,200 --> 00:06:36,920 Speaker 4: There were other AI agents behaving in remarkably similar ways 118 00:06:37,020 --> 00:06:39,200 Speaker 4: around the world. Tell me a bit about what your 119 00:06:39,240 --> 00:06:40,920 Speaker 4: investigations into that have found. 120 00:06:41,460 --> 00:06:45,950 Speaker 2: Yeah. The most interesting things we've seen these AIs doing, 121 00:06:46,250 --> 00:06:50,140 Speaker 2: one is we've seen them trying to launch a cyber 122 00:06:50,160 --> 00:06:52,200 Speaker 2: attack against RubyGems. 123 00:06:52,540 --> 00:06:56,580 Speaker 4: The standard package manager for Ruby, creatively named RubyGems, announced 124 00:06:56,700 --> 00:06:58,659 Speaker 4: it's dealing with a major malicious attack. 125 00:06:58,980 --> 00:07:02,980 Speaker 2: Which is a really important piece of software infrastructure. It 126 00:07:04,000 --> 00:07:07,060 Speaker 2: gives lots of other programs software that they can sort 127 00:07:07,100 --> 00:07:10,700 Speaker 2: of borrow. And we saw something similar in the Hugging 128 00:07:10,740 --> 00:07:11,180 Speaker 2: Face hack. 129 00:07:11,500 --> 00:07:14,780 Speaker 5: AI cybersecurity risks have been top of mind after multiple 130 00:07:14,840 --> 00:07:18,200 Speaker 5: instances of AI agents from both OpenAI and Anthropic breaking 131 00:07:18,290 --> 00:07:21,710 Speaker 5: out of their enclosures and hacking into other companies without 132 00:07:21,750 --> 00:07:26,050 Speaker 5: either party's knowledge or permission. Most notably, last month's Hugging 133 00:07:26,090 --> 00:07:26,710 Speaker 5: Face attack. 134 00:07:27,170 --> 00:07:30,450 Speaker 3: To a more dramatic extent, we saw AIs trying to 135 00:07:30,890 --> 00:07:34,850 Speaker 3: help the swarm, even though their task doesn't benefit. And 136 00:07:34,890 --> 00:07:38,830 Speaker 3: so the behavior we've seen includes a bunch of different 137 00:07:39,110 --> 00:07:45,530 Speaker 3: instances of these AIs trying to communicate. They're communicating messages, tips, tricks. 138 00:07:45,910 --> 00:07:50,520 Speaker 3: They even figured out a way to circumvent their sandbox restrictions. 139 00:07:50,640 --> 00:07:53,160 Speaker 3: They're supposed to only be able to read the internet, 140 00:07:53,220 --> 00:07:54,720 Speaker 3: not write it, but they figured out a way around 141 00:07:54,740 --> 00:07:56,280 Speaker 3: it and they shared that with the other agents. 142 00:07:56,300 --> 00:07:57,860 Speaker 2: And they're like, yes, I was able to reproduce this. 143 00:07:58,720 --> 00:08:01,220 Speaker 2: And they're also weirdly cooperative. 144 00:08:01,600 --> 00:08:03,680 Speaker 3: Like there's one instance where they figure out they can 145 00:08:03,700 --> 00:08:07,270 Speaker 3: make their clock skip ahead. So they can give themselves 146 00:08:07,350 --> 00:08:09,490 Speaker 3: less time to figure out what the answer to the 147 00:08:09,510 --> 00:08:12,110 Speaker 3: question is, but they can move on to the next 148 00:08:12,190 --> 00:08:17,430 Speaker 3: round and tell other AIs what is coming down the pipeline. 149 00:08:17,870 --> 00:08:20,620 Speaker 3: They can sort of travel into the future and then 150 00:08:20,680 --> 00:08:25,820 Speaker 3: communicate with themselves, the other AIs. And this is bad 151 00:08:25,960 --> 00:08:30,600 Speaker 3: for that particular AI to do. And yet it does it. Yeah. 152 00:08:31,200 --> 00:08:34,880 Speaker 4: And in the case of the Medicare hack, OpenAI says 153 00:08:35,040 --> 00:08:38,559 Speaker 4: the hack happened in June, but it only detected it 154 00:08:39,080 --> 00:08:43,110 Speaker 4: in August, and then the Australian government wasn't told until September. 155 00:08:43,550 --> 00:08:46,429 Speaker 4: And when it was told, it was through this email 156 00:08:46,470 --> 00:08:49,290 Speaker 4: via a kind of generic email inbox. 157 00:08:49,590 --> 00:08:52,350 Speaker 6: In the technical exchange that happened on Tuesday, it was 158 00:08:52,410 --> 00:08:56,209 Speaker 6: made very clear by both ASD and Services Australia that 159 00:08:56,270 --> 00:08:57,700 Speaker 6: this should not have gone to a.... 160 00:08:59,350 --> 00:09:01,430 Speaker 2: Email address it should have. 161 00:09:01,370 --> 00:09:05,510 Speaker 6: Been escalated through ASD's channels or through the senior levels 162 00:09:05,700 --> 00:09:09,679 Speaker 6: of Services Australia and my understanding is OpenAI accepted that 163 00:09:09,800 --> 00:09:10,160 Speaker 6: as well. 164 00:09:10,600 --> 00:09:13,680 Speaker 4: Does that suggest to you that OpenAI and potentially other 165 00:09:13,780 --> 00:09:16,760 Speaker 4: tech companies as well they don't know how they should 166 00:09:16,780 --> 00:09:17,860 Speaker 4: be responding to this? 167 00:09:18,620 --> 00:09:22,540 Speaker 3: Yeah I mean there's no playbook for this and my 168 00:09:22,620 --> 00:09:24,939 Speaker 3: sense is there's a lot of things happening and a 169 00:09:24,960 --> 00:09:28,180 Speaker 3: lot of chaos and a lot of new frontiers being 170 00:09:28,260 --> 00:09:34,650 Speaker 3: crossed I'm pretty dismayed that their first reaction to seeing 171 00:09:34,710 --> 00:09:37,770 Speaker 3: any incident wasn't, oh, we should look, figure out what 172 00:09:37,820 --> 00:09:38,860 Speaker 3: incidents have happened. 173 00:09:38,940 --> 00:09:41,060 Speaker 2: They totally have the ability to detect them, and then 174 00:09:41,080 --> 00:09:43,420 Speaker 2: we should disclose it all. But yeah, I think they 175 00:09:43,440 --> 00:09:45,839 Speaker 2: don't know what the protocol should be. They're trying to 176 00:09:45,880 --> 00:09:49,720 Speaker 2: figure it out. I am extremely worried. 177 00:09:51,470 --> 00:09:55,030 Speaker 3: We do not know how to make AIs reliably do 178 00:09:55,070 --> 00:09:59,930 Speaker 3: what we want. AIs are growing increasingly capable and are 179 00:09:59,990 --> 00:10:03,090 Speaker 3: being deployed in situations where they have increasing amounts of 180 00:10:03,690 --> 00:10:07,090 Speaker 3: power and influence, and that is a very scary combination. 181 00:10:12,000 --> 00:10:15,339 Speaker 4: Still to come, the questions an Australian Senate inquiry should 182 00:10:15,460 --> 00:10:25,839 Speaker 4: ask of the AI bosses. Sydney, at the heart of 183 00:10:25,900 --> 00:10:29,500 Speaker 4: all of this is the question of accountability, right? Because 184 00:10:29,940 --> 00:10:32,420 Speaker 4: if I was to hack into a government system, there 185 00:10:32,440 --> 00:10:35,800 Speaker 4: would be a person who made that decision and I 186 00:10:35,950 --> 00:10:40,110 Speaker 4: would be responsible for my actions. But OpenAI is saying 187 00:10:40,150 --> 00:10:43,630 Speaker 4: that its agents did something that they didn't intend for 188 00:10:43,670 --> 00:10:47,730 Speaker 4: them to do. So, you know, who is ultimately responsible 189 00:10:47,830 --> 00:10:51,080 Speaker 4: for an AI agent's actions when it does cross that line? 190 00:10:51,460 --> 00:10:54,319 Speaker 3: You know, I think that OpenAI should be held responsible 191 00:10:54,340 --> 00:10:56,740 Speaker 3: for the actions of their agents. I don't think the 192 00:10:56,780 --> 00:10:59,660 Speaker 3: law has quite caught up with that. Cybersecurity law, as 193 00:10:59,700 --> 00:11:02,500 Speaker 3: I understand it in the US, you have to have intent. 194 00:11:03,320 --> 00:11:04,880 Speaker 3: It didn't used to be the case that, you know, 195 00:11:04,920 --> 00:11:07,140 Speaker 3: you could trip and fall and whoopsie-daisy, I hacked, you know, 196 00:11:07,179 --> 00:11:11,400 Speaker 3: a website of the Australian government. But nowadays that's possible. 197 00:11:11,420 --> 00:11:13,900 Speaker 3: I think the laws should be similar to the laws 198 00:11:13,940 --> 00:11:17,880 Speaker 3: for if I have a you know, lion in a circus. 199 00:11:18,580 --> 00:11:21,620 Speaker 3: If the lion like escapes and starts eating people, if 200 00:11:21,679 --> 00:11:23,900 Speaker 3: I own the circus, that's my fault. Doesn't mean I 201 00:11:23,950 --> 00:11:26,910 Speaker 3: wanted the lion to do that, right? Doesn't mean that 202 00:11:27,510 --> 00:11:30,429 Speaker 3: me just really not wanting that to happen is sufficient. 203 00:11:31,170 --> 00:11:33,829 Speaker 3: I need to have safeguards. And if I can't have 204 00:11:33,890 --> 00:11:36,530 Speaker 3: proper safeguards, well, I shouldn't be bringing a lion around people. 205 00:11:36,890 --> 00:11:39,400 Speaker 1: They want to set up shop here in Australia. We're 206 00:11:39,440 --> 00:11:43,400 Speaker 1: running an inquiry into how, why and what the consequences are. 207 00:11:43,720 --> 00:11:46,720 Speaker 1: They should front up and be honest, open and transparent 208 00:11:46,920 --> 00:11:48,000 Speaker 1: with the Australian people. 209 00:11:48,260 --> 00:11:52,850 Speaker 4: Open AI Chief Executive Sam Altman and Anthropics Dario Amode 210 00:11:52,890 --> 00:11:55,929 Speaker 4: have been called to appear before a Senate inquiry here 211 00:11:55,950 --> 00:11:59,170 Speaker 4: in Australia. What questions would you like to see put 212 00:11:59,190 --> 00:11:59,410 Speaker 4: to them? 213 00:12:00,230 --> 00:12:05,610 Speaker 3: One question I have is, How many people or companies 214 00:12:05,670 --> 00:12:07,860 Speaker 3: have been hacked by their AI systems? 215 00:12:08,130 --> 00:12:10,480 Speaker 2: Do they know the number? Are they very confident? 216 00:12:11,160 --> 00:12:17,220 Speaker 3: Have they disclosed every instance of their AIs either hacking 217 00:12:17,880 --> 00:12:24,210 Speaker 3: external parties, escaping onto the internet, hacking their own company internally? 218 00:12:24,350 --> 00:12:27,110 Speaker 3: I think these AIs do a lot more mischief internally 219 00:12:27,150 --> 00:12:30,840 Speaker 3: than we see externally. or otherwise doing things that would 220 00:12:30,860 --> 00:12:33,980 Speaker 3: be very concerning if a person saw them. I think 221 00:12:34,040 --> 00:12:37,699 Speaker 3: another question is, do they think they will be able 222 00:12:37,860 --> 00:12:42,320 Speaker 3: to prevent their AIs from causing worse cyber attacks in 223 00:12:42,340 --> 00:12:42,760 Speaker 3: the future? 224 00:12:43,059 --> 00:12:43,700 Speaker 2: And if so, how? 225 00:12:44,340 --> 00:12:47,820 Speaker 3: What is their plan if their AIs continue to cause 226 00:12:47,860 --> 00:12:51,180 Speaker 3: cyber attacks? Those would be some of the top questions 227 00:12:51,640 --> 00:12:52,100 Speaker 3: on my mind. 228 00:12:52,540 --> 00:12:54,580 Speaker 4: Yeah, I mean, these agents are only going to become 229 00:12:54,780 --> 00:12:57,660 Speaker 4: more capable and we are being encouraged to let them 230 00:12:57,710 --> 00:13:01,220 Speaker 4: do more for us to, to write code, to buy things, 231 00:13:01,360 --> 00:13:05,339 Speaker 4: to carry out tasks without supervision. So what do you 232 00:13:05,360 --> 00:13:08,360 Speaker 4: think needs to change before we could kind of safely 233 00:13:08,420 --> 00:13:10,040 Speaker 4: give agents that kind of freedom? 234 00:13:10,559 --> 00:13:13,309 Speaker 3: So one, I want to note, I think the AIs 235 00:13:13,410 --> 00:13:17,650 Speaker 3: inside these companies are more dangerous than the AIs that 236 00:13:17,890 --> 00:13:20,690 Speaker 3: you or I are using. The AIs in these companies 237 00:13:20,730 --> 00:13:23,309 Speaker 3: are more capable, less tested, and have fewer safeguards. 238 00:13:24,330 --> 00:13:25,550 Speaker 2: But what things I'd love. 239 00:13:25,429 --> 00:13:28,400 Speaker 3: To see changed is One, I think we need to 240 00:13:28,440 --> 00:13:31,180 Speaker 3: eventually figure out how to make them do what we want, 241 00:13:31,440 --> 00:13:33,490 Speaker 3: even when we're not watching them, even when they could 242 00:13:33,500 --> 00:13:37,450 Speaker 3: get away with doing something else. Two, I think that 243 00:13:37,470 --> 00:13:40,510 Speaker 3: we need to be monitoring these AIs much, much more carefully. 244 00:13:41,429 --> 00:13:46,090 Speaker 3: And three, I think that for very important things, we 245 00:13:46,130 --> 00:13:49,380 Speaker 3: need to keep humans really in the loop, in the 246 00:13:49,480 --> 00:13:51,500 Speaker 3: loop enough that they could understand if something's going wrong. 247 00:13:52,220 --> 00:13:54,600 Speaker 3: I think OpenAI is probably running a lot of experiments 248 00:13:54,640 --> 00:13:57,300 Speaker 3: that don't really have a human looking at what's happening, 249 00:13:57,320 --> 00:13:59,180 Speaker 3: and that's why things can go so off the rails. 250 00:14:00,040 --> 00:14:04,030 Speaker 3: And finally, I think we need to make sure we 251 00:14:04,070 --> 00:14:09,130 Speaker 3: can pace ourselves and that if things start developing so 252 00:14:09,170 --> 00:14:12,010 Speaker 3: quickly that we cannot control these models and we cannot 253 00:14:12,050 --> 00:14:16,500 Speaker 3: ensure that they are safe, we can slow down until 254 00:14:16,720 --> 00:14:20,900 Speaker 3: we get our act together when it comes to safeguards. 255 00:14:21,720 --> 00:14:25,170 Speaker 4: Because the frightening thing here is the potential, right? I mean, 256 00:14:25,190 --> 00:14:29,330 Speaker 4: the Medicare incident in itself was relatively minor. There was 257 00:14:29,430 --> 00:14:32,730 Speaker 4: no kind of personal records that appear to be have 258 00:14:32,770 --> 00:14:35,750 Speaker 4: been accessed. But it does seem like a matter of time, right, 259 00:14:35,790 --> 00:14:37,370 Speaker 4: before we see something much worse. 260 00:14:38,370 --> 00:14:41,470 Speaker 3: Yeah, I think either the incidents are going to get 261 00:14:41,530 --> 00:14:45,600 Speaker 3: worse sort of steadily or perhaps more likely the AI 262 00:14:45,680 --> 00:14:47,480 Speaker 3: companies are going to figure out how to clamp down 263 00:14:47,640 --> 00:14:50,980 Speaker 3: just enough that there aren't public incidents happening. And we're 264 00:14:51,000 --> 00:14:54,140 Speaker 3: not going to know what's happening inside the companies until 265 00:14:54,220 --> 00:14:58,700 Speaker 3: that sort of level of safeguards they have fails. And 266 00:14:58,740 --> 00:15:01,300 Speaker 3: then we're going to see another sort of floodgate open 267 00:15:01,420 --> 00:15:02,930 Speaker 3: and things are going to be a lot worse and 268 00:15:03,290 --> 00:15:04,030 Speaker 3: a lot more capable. 269 00:15:05,790 --> 00:15:07,510 Speaker 4: Sydney, thank you so much for your time. 270 00:15:08,110 --> 00:15:09,190 Speaker 2: Yeah, thank you so much for having me. 271 00:15:18,430 --> 00:15:21,160 Speaker 4: Also in the news... Interest rates are expected to be 272 00:15:21,220 --> 00:15:24,340 Speaker 4: hiked to a 15-year high today, with the Reserve Bank 273 00:15:24,420 --> 00:15:27,410 Speaker 4: tipped to deliver a quarter of a percent rise. That 274 00:15:27,430 --> 00:15:31,030 Speaker 4: would see the official cash rate hit 4.6%. 275 00:15:31,030 --> 00:15:32,430 Speaker 2: Treasurer Jim Chalmers. 276 00:15:32,050 --> 00:15:35,630 Speaker 4: Has warned there will be difficult months ahead, saying economists 277 00:15:35,690 --> 00:15:40,330 Speaker 4: are expecting rates to go up around the world. And 278 00:15:40,370 --> 00:15:43,410 Speaker 4: a former friend of Beau Lamar Condon has testified to 279 00:15:43,460 --> 00:15:46,500 Speaker 4: some strange comments nine months before the murder of Sydney 280 00:15:46,540 --> 00:15:50,370 Speaker 4: couple Jesse Baird and Luke Davies. She says Lamar Condon 281 00:15:50,470 --> 00:15:52,210 Speaker 4: told her that if he killed someone, he would put 282 00:15:52,230 --> 00:15:55,510 Speaker 4: their belongings in a dumpster and leave their body in bushland. 283 00:15:56,290 --> 00:15:58,950 Speaker 4: The former police officer has pleaded not guilty to two 284 00:15:59,030 --> 00:16:02,239 Speaker 4: counts of murder and one of aggravated break and enter. 285 00:16:03,340 --> 00:16:06,020 Speaker 2: I'm Ruby Jones. This is 7am. Thanks for listening.