'Terrifying warning sign': Anthropic delays AI model over security concerns

[0:00] Anthropic has decided not to release its latest AI model called Clawed Mythos to the full public. [0:06] Now, this version is apparently so advanced it can find vulnerabilities in a huge variety of software applications, [0:12] which is both a great advancement potentially for improving cybersecurity [0:16] and also a huge danger if used by criminals or others to hack into systems. [0:21] For now, Anthropic is only making this system available to some of the biggest tech companies [0:25] so they can test it and improve their own systems. [0:29] The fear, of course, if it falls into the wrong hand, cyber criminals, spies, the fallout could be catastrophic. [0:36] My next guest, Tom Friedman, lays out those concerns in his latest New York Times op-ed with the headline, [0:41] Anthropic's Restraint is a Terrifying Warning Sign. [0:44] He's also the author of the best-selling book, among many others, From Beirut to Jerusalem. [0:49] So, Tom, you write this is potentially as fundamental and significant a turning point [0:53] as was the emergence of mutually assured destruction and the need for nuclear non-proliferation. [1:00] That's a huge statement. [1:01] How do you say that and what is your biggest concern? [1:06] Well, you know, Anderson, our economy now, all our biggest systems, our water systems, [1:11] our airlines, our airports, our transportation and telecommunication systems, [1:16] all run on software and operating systems, as do those of every other major economy in the world [1:22] that we are interlinked with. [1:25] And if, basically, we now have a software that not only is fantastically good at writing new code, [1:34] but it turns out is fantastically good at finding bugs in your own code or other people's codes, [1:43] that tool, that power now can be used all over the world. [1:48] It will be in the hands of everyone. [1:50] Imagine a world where everyone had a nuclear bazooka, basically. [1:53] And because it's so easy, a threat, a cyber threat that used to be confined to intelligence organizations, [2:00] big companies, which are very hard and very expensive to do, [2:04] is to become really cheap and really easy. [2:07] And that's why Anthropic took off. [2:09] Yeah. [2:09] And Anthropic, what they've actually said, what they've disclosed is that, quote, [2:13] they say the vulnerabilities in this program, it has spotted, [2:18] have in some cases survived decades of human review and millions of automated security tests. [2:23] I mean, that's mind-blowing that all the current, you know, [2:27] human review, humans looking at it and all the current, you know, [2:31] AI security tests that they have run programs through and software through, [2:35] all of it still had bugs that were detected by this new powerful AI. [2:42] Well, you know, this is a point that my, you know, [2:44] technology tutor and partner in this column, Craig Mundy, [2:47] keeps making that this, these AI capabilities, [2:51] they're coming so much faster than people realize. [2:55] And that we aren't, as Craig says, birthing just a new tool. [3:00] Anderson, we're birthing a new species. [3:03] It's not carbon-based like we are, but silicon-based. [3:06] But it is a new species that we're going to have to learn to control [3:09] and collaborate before it makes us its pet. [3:14] So this is the front end of a really big problem. [3:17] Well, regulation, control is not a word that this White House is very interested in, [3:23] in terms of artificial intelligence. [3:26] Anthropic is probably one of the few AI companies. [3:29] Dario Amadei, who left OpenAI years ago with his sister and several others to form Anthropic, [3:35] they're probably the most kind of safety forward. [3:38] Do you think that there will be, given now this development and others, [3:44] some sort of move to, you know, really look at kind of regulation or even government to government? [3:53] I mean, even with some of America's adversaries, [3:55] you've written about the idea of the U.S. and China kind of getting together on this. [4:00] Well, you know, I think, Anderson, this is the front end of a fundamentally new problem [4:04] we have as a species, okay? [4:07] So I started out in journalism, and you started out when the world was connected. [4:11] During our tenure with the Internet, the world became interconnected. [4:15] What's happened today is the world has become interdependent, okay? [4:20] As my teacher, Dove Seidman, likes to say, [4:22] interdependence is no longer our choice, it's our condition. [4:26] Now, we are going to rise together or we're going to fall together. [4:30] But, baby, whatever we're doing, we're doing it together. [4:33] Now, we can discover that early or we can discover that late, but we will discover that. [4:38] It's true about climate, but it's also true now about cyber tech and all of these communication systems. [4:45] No company alone can solve this problem. [4:48] Therefore, no country alone can solve this problem. [4:51] And therefore, sooner or later, the two tech superpowers, the U.S. and China, [4:56] are going to have to sit down and work out this problem together, [4:59] learn how to compete and cooperate on AI, because we are interdependent. [5:06] That is our condition. [5:08] It's not a choice. [5:09] Trump can learn that early. [5:11] He can learn that late. [5:12] But he will learn that. [5:14] If you have, though, these powerful AI weapons at your disposal, [5:18] and if it's an American company that has it, [5:21] if there's an AI arms race, which is how many people describe what is happening around the world, [5:27] isn't there a kind of a push at some point or pressure at some point for an American company [5:34] to in some way use this for the benefit not only of their own company, [5:38] but for a country's benefit? [5:41] The idea of some sort of China-U.S. alliance on this seems far, far, difficult to imagine right now. [5:49] Hard to imagine, but, you know, when I say it to people, they say, [5:54] God, that's incredibly naive. [5:56] I say, no, no, no, no. [5:57] What's naive is thinking we're going to be okay if we don't do that, [6:02] because there's no way both China and America will be hugely vulnerable [6:06] to criminals within their own societies when they have these tools, [6:10] much more vulnerable than they're going to be to any threat the other would pose. [6:15] So, again, we're in the middle of a giant meta transition. [6:20] You know, I've been working on a new book on this. [6:22] We have become godlike as a species. [6:24] We've just created this super intelligent being. [6:27] There's just one problem. [6:28] We've become godlike without the Ten Commandments. [6:32] And, you know, we're going to sit down and together with the biggest companies [6:37] and our biggest rivals and write a new Ten Commandments, [6:41] Ten for us and ten for the new species we've just birthed. [6:45] OpenAI is calling for governments to implement common sense AI regulation. [6:51] On the substance of the proposal itself, first, what do you make of its approach [6:57] and how do you think lawmakers will respond? [7:00] Because Congress, historically, is fairly slow moving when it comes to new technology. [7:07] Oh, yeah. [7:07] Slow moving is one way to put it. [7:09] At sea is another way to put it, right? [7:11] They have been totally incapable of wrapping their arms around what this may represent. [7:18] I mean, you'll remember, Boris, Andrew Yang, who once upon a time ran for president [7:21] on the concept that we would all need to somehow draw from a universal pool of money [7:26] because AI was going to wipe out all jobs. [7:28] You know, we laughed at him back in the day. [7:29] That guy must be drinking like a fish these days. [7:32] Because, I mean, it is absolutely the case here that you now have this company saying openly [7:37] that its technology is going to create incredible upheaval, [7:41] so much so that what it's proposing here is pretty interesting, Boris. [7:44] I mean, it comes up with a number of things. [7:45] It suggests, for instance, that we should all be working a four-day work week with no loss of pay, [7:50] that we should have a sovereign wealth fund, essentially a kind of general fund [7:55] that would help bring some dividends back to workers who currently aren't seeing anything. [7:58] And this is a big one which breaks from other big tech companies, [8:02] which is the idea that the government needs to stop depending on labor payroll [8:07] as a source of income tax and instead create a tax entirely on capital gains and corporate profits [8:15] because basically what they're saying here is that you and I and our paycheck [8:18] isn't going to be enough to support the government anymore. [8:20] All that money is going to be vacuumed up into these larger corporations. [8:24] So there's some good and interesting ideas here. [8:26] But as you brought up so rightly, you know, the timing here, [8:29] distracting, I think, from a big and embarrassing investigation on the part of the New Yorker [8:33] is also worth mentioning here, Boris. [8:35] One more question on the substance of the argument for regulation. [8:41] Altman talks about superintelligence, the point at which AI becomes so sophisticated [8:47] that it would out-compete or out-smart even humans using AI. [8:52] How close do you think we are to that point? [8:54] So I am not personally convinced that we are anywhere near that. [8:59] I think that it is some, you know, the technical people that I speak to, [9:02] including people who used to be at OpenAI and Anthropic and others of these big companies, [9:07] basically say we're not on the path to that right now, [9:10] that the kind of very smart parrot that we currently use to write our knock-knock jokes [9:15] and our wedding vows is not going to put us toward, you know, [9:18] an omniscient brain that's going to run all things. [9:20] But we are already on a path where job disruption is real, [9:24] where human anxiety around AI is real. [9:27] And so what's interesting here is to see a company that is really at the center of that disruption [9:32] calling this kind of a new deal for AI, trying to invoke the Roosevelt era. [9:39] But of course, that was an era grappling with the depression, [9:42] and this is the company that is essentially saying we may cause a depression here, [9:46] and so we're going to need a national response to deal with it. [9:49] There's a circular logic to this kind of statesmanship on the part of a company [9:55] that really seems to be driving the change it's worried about Boris.