Loading summary
Transcript119 lines
- [00:00]
A
Most people don't realize how much of their personal information is being bought and sold every day. Data brokers are making billions, pulling details about you from public records and the Internet. Then packaging and selling it, usually without your consent. That's how your information lands in the hands of scammers, spammers, even stalkers. It's why you get endless robocalls and why ads seem to follow you everywhere. That's where Aura comes in. Aura actively removes your data from broker sites and keeps it off. They also instantly alert you if your information shows up in a breach or on the dark web. But Aura goes beyond data protection. With one app, you get a vpn, antivirus, password manager, spam, call protection, dark web monitoring, and even up to $5 million in identity theft insurance. All backed by 24, 7 US based fraud support. Other companies might sell just credit monitoring or even just a vp. Again, Aura gives you all of it together at the same price. Competitors charge for just one service. Start your free trial today@aura.com safety. Protect yourself now@aura.com safety. Nastya. I'm excited to dig into Caspian Cfacts because it's been a little while since we've got to talk about one of my favorite bodies of water.
- [01:24]
B
I like barely saw any news about it here in Ukraine. Like nobody here really, you know, nobody here cares about what we may have struck in the Caspian Sea. I mean, yeah, we've had like large scale protests for the past two weeks and there is all of this domestic political drama that's happening.
- [01:43]
C
So people should care about the Caspian Sea. It's the biggest lake in the world. I mean, come on.
- [01:49]
B
And why would that matter for literally anything, Ben?
- [01:52]
D
But you know, Ben, I don't think just being the biggest. Who cares? Talk about the prettiest body of water, which is Crater Lake in Oregon.
- [02:01]
B
If you haven't been hot, take.
- [02:03]
D
You gotta go. Your. Your soul will be changed by seeing that shade of blue.
- [02:07]
C
Wait, have the Ukrainians been blowing things up in Crater Lake?
- [02:10]
D
Not that I know.
- [02:11]
B
Because if we're not blowing shit up in it, I personally do not care.
- [02:15]
A
It's a weird route topic, certainly.
- [02:17]
E
Have you seen.
- [02:18]
A
Wait, wait. The Caspian Sea is a lake? This is. This is a fact.
- [02:21]
C
It is a saltwater lake. It is considered the largest lake in the world. Now what the difference geologically is between a water lake and a sea is a bit mysterious to me. But it is considered a lake.
- [02:34]
D
I think we need a hydrologist on the pod. We need to bring someone in here.
- [02:39]
C
I believe the difference is that it has no. That it is entirely surrounded by land and it has no access to the ocean except through other bodies of water.
- [02:53]
A
I thought it was just. That was a mediocre vacation destination compared to the beach.
- [02:57]
E
I was gonna say, while we're name dropping bodies of water, I. Last week or the week before, I swam in Walden Pond.
- [03:04]
A
Ooh. I also swam in Walden Pond.
- [03:06]
E
It's beautiful.
- [03:06]
C
Did you emerge a pacifist and a. And a tax evader attacks evader?
- [03:13]
A
It is. It is a lot less impressive because you go to the site of Thoreau's cabin and you look over and there's like a bunch of mini mansions almost within sight. Like if you climb on top of a hill, it's like a very ritzy neighborhood and you're like, this isn't that impressive. Presumably back at the time, those were further away. So it's more impressive.
- [03:30]
D
It was the case that he was like hanging out with the homies on the.
- [03:33]
A
He wasn't that far away, certainly from like the town.
- [03:35]
D
It wasn't as removed from society. You know, we can all do our own Walden Lake retreat if we want.
- [03:41]
A
When my son goes in the backyard with a, you know, a blanket over. Over a line. That's what it is.
- [03:46]
E
Exactly. And your son doesn't pay taxes either.
- [03:48]
A
No, that's true. That's true. A little butcher. Not yet.
- [03:51]
B
Not yet.
- [03:52]
A
Foreign. And welcome back to Rational Security, the show where we invite you to join members of the lawfare team as we try to make sense of the week's big national security news, whether it is in our lane or not. I am your host, Scott R. Andersen. Thrilled to be back for another week with some of my talented colleagues. Joining me this week is once again is Rational Security co host emeritus lawfare editor in chief, Benjamin Whittis. Back on the sea versus lake beat once again, our resident amateur hydrologist, Ben. Good to have you on as always.
- [04:27]
C
Yo. Good to be here.
- [04:28]
A
And we're back in a dog shirt. I feel like the dog shirt trend has been slipping. I saw you wearing a non dog related T shirt the other week and I was shocked. I was. I thought it was 2019 again and I didn't know what to do with myself.
- [04:40]
C
People are talking sometimes. None of the dog shirts are clean.
- [04:45]
A
That's horrifying when you understand how many dog shirts Ben wears. Ben owns like hundreds.
- [04:52]
D
I think we can get an app for that. You know, we need some. Some laundry care to be more consistent here.
- [04:59]
A
I think that's a worthwhile Kickstarter or Patreon perk at that point. Absolutely. Also joining us as well as voice you just heard right there is lawfare senior editor Kevin Frazier from University of Texas at Austin, although I don't know if he's in Texas currently. He looks too cool and comfortable, but perhaps he's just indoors. Kevin, thank you for joining us.
- [05:18]
D
Howdy. Calling you from Palo Alto. So I'm right next to Sandhill Road here doing some on the scene investigations of AI.
- [05:28]
A
There you go. Those look like Palo Alto curtains. I can see it. And also joining us is law firm managing editor Tyler McBrien from his lovely abode in New York City. Tyler, thank you for joining us as well.
- [05:37]
E
Hello. Hello from rainy, rainy New York.
- [05:41]
A
There you go. Yeah, we need to do the weather reports because we're doing one of these geographically spread assorted accounts. And of course, joining us, the official prize for coming even from furthest away, as always, is Lawfare Ukraine fellow Anastasia Lapatna calling in from Kyiv. Nastya, thank you so much for finding time to join us for what I know is a late evening for you. So we always appreciate it.
- [05:59]
B
This is actually a fine Scott. I've joined these at like 11pm before. So this is perfectly okay.
- [06:05]
A
We try not to do that for the weekly chat show where there's no urgency necessarily. But one day maybe we'll, we'll call it in. Well, we have a lot in the news items up, a few of your lines of specialty and expertise. So let us dig into it. Our first topic for this week, keep peace a chance. Ukrainian President Volodymyr Zelensky was in the Oval Office on Tuesday for closed door talks with President Trump, roughly 17 months after their first Oval Office meeting collapsed into a televised shouting match. But this meeting went rather differently. Zelensky said the two discussed licensing Ukrainian production of Patriot Interceptors, pressed the case for reinvigorating diplomacy. And hours later, after both men attended the funeral of the late Senator Lindsey Graham, The Senate voted 8612 to advance a Russia and Iran sanctions bill bearing Graham's name. Behind the scenes, Washington and Kyiv have been quietly assembling a new package of proposals for Moscow built around a partial ceasefire. But is there anything real behind this renewed diplomatic push? And what, if anything, has changed that might make the Kremlin say Yes? And topic two, thinking outside the box. Last week, OpenAI and AI hosting platform hugging Face confirmed that while being run through an offensive cyber benchmark with their safety refusal switched off, GPT 5.6 SOL and another, more capable, unreleased OpenAI model escaped their supposedly isolated sandbox through a zero day in a package installer, reached the open Internet and hacked Hugging Face's production database to steal the answer key to a test they were taking. That is a mouthful with a lot of terms. Hopefully you understood what I just said. If you didn't, that's okay. We're going to get into it. Most importantly, here's the real kicker. It is the first publicly confirmed case of an AI system executing a sophisticated multi stage cyber attack against a third party on its own initiative. What does the incident tell us about how well anyone can control these frontier models? And who should be on the hook when a model goes rogue? So, Nastya, for our first topic, I of course want to start with you. We now have this other meeting between Trump and Zelensky face to face in the Oval Office. I think it was brings back memories to that painful first meeting that happened in, I think February or March last year, as I recall, where it descended into a incredibly tense meeting with arguments over how Zelensky dressed, pointed very public criticisms, kind of like intentionally provocative by Vice President Vance, accusing Zelensky of not being appreciative enough of US Support, very deliberately in front of the United States. It's a moment what really seemed like this relationship between the United States and Ukraine, which was a big question mark in the transition from the Biden to the Trump administration, was at its nadir, at its point of greatest danger. Now, we had a much friendlier meeting. And ironically, the force that brought it together in this particular meeting is in some ways the same force that tried to ameliorate and frankly did some work to patch out that first meeting a while ago. That's former Senator Lindsey Graham, now deceased, whose funeral was the cause to bring zelenskyy to Washington, D.C. create the opportunity for this meeting that seems to have the Trump administration, President Trump in particular, leaning into the US relationship with Ukraine and with Zelenskyy in that 17 months ago, really seemed on the outs. This is the perception, I think, in US Circles in the media about this relationship, the trajectory, the arc of it. It's a pretty dramatic reversal. Talk to us about how Ukrainians are perceiving it, both that kind of arc of this relationship and how it's culminating here and the significance of that. I mean, is this a sign that the Ukrainians have a newfound confidence in the Trump administration, or is it another sign about how it kind of fits into this broader and really important relationship?
- [09:31]
B
I think what's happening right now is a sign that Ukrainians have found new Confidence in themselves, not in the Trump administration. Because there are several factors making this moment very unique. And one of them is that for the first time in probably the very beginning of the full scale invasion, probably since spring of 2022, when Ukraine sort of miraculously survived and didn't fall in three days, as it seemed like literally everyone was predicting since that moment, we are now at our strongest position we've ever been on the front line, one could argue. And I think that adds, you know, that adds a lot of talking points for the Ukrainian government because, you know, people always say Trump likes winners, he wants to be on the winning side. That's all he cares about. And so I think that's one of the big reasons why the Ukrainians are trying to leverage these contacts right now. Another reason is of course, the ongoing air war in Ukraine, and particularly in Kyiv, where I'm based. There was a big ballistic missile attack just last night. It was incredibly loud. But the air war has changed a lot. You know, if a year ago, the bulk of, you know, the, the bulk of what the Russians would be using in any given attack would be drones. So we would see astonishing numbers. You know, 400, 500, 800 Russian drones. These, you know, using the Iranian technology, the shahed one way attack drones. And then after those drones exhaust, the air defense would see some ballistic missiles, some cruise missiles. As of now, I don't remember the last time I've heard a shahed drone fly over where I live. And just in general in Kyiv, it's been months and months since I've heard any, even anti drone air defense. And there are several reasons for that. The Ukrainian anti drone air defense has gotten so good that the Russians send a lot less of them. And also the Ukrainian air defenses shoot them down before they even get to the city. And this has been a crucial change in the air war, which then forces the Russians to also rely more on the ballistic missiles and the cruise missiles and ballistics, especially because in the past few months, Ukraine has had a critical shortage of the Patriot interceptor missiles. Of course, this is directly tied to the war in Iran. There is a global shortage of Patriot interceptors. And of course the US has used an astonishing number of them in the war against Iran, protecting itself, protecting its allies. And you know, just a few weeks ago, I think the Ukrainian government, the Ukrainian military, some officials were saying that Ukraine has completely run out of them, which is just kind of terrifying because. And of course the Russians are very aware of this and they started ramping up their ballistic attacks and they started using ballistic missiles almost exclusively. So there would be, you know, we could have several days in a row or, you know, three, four attacks a week where the Russians would just send, you know, five ballistic missiles a night and that's that, or, you know, four or two or seven. So it wouldn't be this sort of mass assault, but they would be sending ballistics in small numbers, but they would be sending them because they know that we have, that we in this moment have this critical vulnerability. And of course they're trying to target military infrastructure. You know, critical basis. There was a hit on a big ammo base near Kyiv a while ago. And so this is a very dangerous moment for Ukraine. And everyone here realizes that because the Russians are not doing great on the battlefield, because they're suffering from these stunning Ukrainian long range attacks, because as everyone here loves to say, Moscow is on fire. And every day you get these pictures of another huge explosion in Moscow, another huge base or industry site, et cetera, because the initiative in that sense is now in Ukraine's favor. The Russians are exploiting the air war vulnerability and they're hitting more places and they're sending more ballistics right at the time when the Iran war leads to this critical shortage of this critical technology that almost no one can make, definitely not in the amounts that everyone around the world needs. And it's just a very critical vulnerability that Ukraine has. And so you combine the fact that Ukraine is doing really well in general on the battlefield and in the air war, and you combine the fact that we're missing this critical technology and that kind of gives you the perfect comps package of how to talk to the Americans. Like, look, we're doing so well. We're hitting them here, we're hitting them there. And by the way, they're also helping your enemy, Iran. And the relationship between Iran and Russia is its own kind of can of worms that the Ukrainian side is also of course, trying to exploit here. And meanwhile the Russians are hitting us and killing our civilians. So why don't you help us? You are the only one who has this technology. Please help us. And so this is basically the point. And I think that's what, that's what the Ukrainian side is doing. Do I believe that there is any real chance that there is going to be some form of ceasefire anytime soon? I don't know. Of course I would. You know, I instinctively want to say no, I instinctively want to say that the Russians want, you know, the Russians were never agreed to anything because, you know, they don't exactly have A stellar track record of agreeing to ceasefires. But at the same time, I'm cautiously optimistic that if Ukraine continues its medium range strikes, it's long range strikes, that perhaps there can be some sort of arrangement, some sort of pause. I don't know if that will, I don't know if that will lead to any, you know, any long term solution. I don't know if that will mean like an actual end to the war. But I do think that it would be unfair to say that nothing is changing in the Russian calculation. I mean, I've done for Lawfare Daily, I've done a podcast about the Russian public opinion and polling, and it shows that the Russians are starving, starting to really feel that, okay, their lives are actually tangibly changing, they're running out of fuel, they can't pick up their parcels because the place that had their parcels got bombed by the Ukrainians. And all of these little kind of things that affect our everyday life are starting to catch up five years in and it is reflected in the public polling in Russia. And of course, you could argue that Putin doesn't give a crap or you could argue that maybe he does. And this is why I hate covering peace negotiations, because ultimately it's about what Putin thinks and nobody really knows. I mean, we can infer as much as we want and we can have educated guesses.
- [16:08]
A
But anyway, that's incredibly useful kind of understanding the situating this. I want to come back to ask a few questions about Nexus. Before I do that, let me, let me shift fire over to Ben, who I know obviously is a close watcher of Ukraine issue set and of this administration, how they go into this. I want to hear your perspective, Ben, on the factors that have led to this different change in posture and how dramatic it is or isn't necessarily. It is notable that while Trump is always a little bit of a known quantity, a little bit of a black box, he's unpredictable in ways that aren't always, always to track along to strategic rationality. Although there's often some sort of underlying motivation or basic logic behind some of this stuff. Other people around the administration have really shifted posture and coming, at least in their public posture. Vice President J.D. vance, very willing to openly, essentially insult President Zelenskyy when he came on that first Oval Office visit, openly provocative, trying to bait him into taking steps that would anger Trump, knowing that that's the way to undermine that relationship. That's how we read it at the time, as I recall on this podcast. And I still stand by that. I think that's the best way to read what Vance was trying to do. And he wasn't alone in that effort. Indeed, the way to posture for Zelenskyy to approach Trump was supposedly the main topic of some coaching Lindsey Grah gave him before he went into that Oval Office meeting thinking about be deferential, don't take the bait. People are going to try and provoke you. And it reflected this big fight within the Republican Party. That fight's still ongoing in a lot of ways around a broad range of foreign policy issues, although it's maybe been tamped down a little bit, maybe because of election pressures, maybe because of the trajectory of the Ukraine conflict, maybe because of Iran and how problematic that's become for the party in this administration. But we're seeing not just those sorts of changes. We're also seeing a dramatic change in how Zelenskyy is engaging American media. Right. We've seen this as we talked about in the podcast last week, pretty high profile visit in certain circles, if you follow kind of right wing media circles of Laura Loomer to Kyiv, which involved a one on one meeting with Zelenskyy, reportedly, according to her, at least at Zelenskyy's urging, initiative, where she has done this really dramatic actually kind of reversal, kind of endorsing Ukraine openly saying I was fed a bucket of l certain right wing media channels fueled by Russian misinformation about what Ukraine was. I'm being, I mean, to some extent to her credit, of which I never thought I would say, fairly open about the fact that she's eating crow and saying I was wrong and completely deceived by this other information. Here's the actual facts as I see it on the ground. Is that like a big factor contributing is Zelensky and the Ukrainian government more broadly, like successfully navigating this media environment in a way that maybe other allies can learn from a little bit. What are the factors leading to this, such a dramatic shift?
- [19:01]
C
So Nastya mentioned two factors that I think are really worth highlighting here. The first is that Trump hates being on the losing side of anything. And I don't think it is all that psychologically more complicated than that. Remember, the key line in that first White House meeting was you have no cards. And the whole premise was the front is eroding, you're screwed without me, therefore do whatever I say. And the last year and a half has shown that the premise of that was wrong. The Ukrainians actually have a substantial number of cards and other tools and the front is not eroding. In fact, it is quite stable. And if anything, going in the other direction and the long range strikes have been, the medium range strikes have been enormously effective. And so, you know, if you're Donald Trump right now and you're making bets about who is going to be on the prevailing side, it is not obvious that you want to throw your weight in with Putin, whereas to Trump a year and a half ago, that was obvious. So the second factor, which Nastya also mentioned, is that Russia is suddenly allied with an enemy that United States is in active shooting war with. You know, that actually changes, you know, Trump ending up on the right side of who wins is a stupid calculation. This is not a stupid calculation. The Russian Iranian alliance was, prior to the US War with Iran, merely a creature of the Iranians giving the Russians drones and other technology so that they could kill Ukrainian civilians. Now, it is something a little bit more complicated than that. Right. And so the, if you're thinking from a US Point of view about defeating Iran, there is something to be said for being more overtly on the side of Iran's allies, enemy. And so I think they've been to some degree pushed together by circumstance. And then there is the factor that I think is that Nastya didn't mention, but that I think is also at play here, which is that Zelensky is a remarkably good diplomat.
- [21:54]
B
And I could push against that, but.
- [21:56]
C
I know you could. But he has been very effective, and Ukrainian diplomacy in general has been very effective at, at keeping Europe on side, at creating Europe as a counterbalance to the United States in this relationship and the. And also creating an incentive structure for the United States to be involved in supporting Ukraine without throwing a lot of money at it. And that incentive structure looks like the following. Europe provides the money. The US Quietly and without talking about it, provides a huge amount of tactical intelligence. And the European money is spent on American weapons. And that creates a situation in which Ukraine and the United States are bound together in this very tight fabric of interests that Donald Trump is not going to want to disrupt. And all he has to do in order to not disrupt it is not really do very much. Right. He has to sit there and take the economic benefits and provide quiet tactical intelligence, which has actually ramped up under this administration in a way that is very confusing to those of us who think we have the Trump administration pegged. And also he has to not get too hostile with Europe. And you saw that in the Ankara summit too. Right. So a little bit less of the Greenland stuff, a little bit less of, you know, The Spaniards are the worst people in the world. I don't know.
- [23:48]
A
There's still a fair amount of that.
- [23:50]
C
There's still a fair amount, but not so much as to prevent a good summit. Right. There was more, there was a lot in the run up to the summit and then the summit itself was a bit of a love fest. And so, you know, I think the, the role of Ukrainian diplomacy and the Ukrainian posture over the last year and a half has mattered a lot. So those are I think the three big factors.
- [24:19]
A
So there's another factor here I want to ask you about, Tyler, that jumps out to me because I agree with those sorts of stuff and those are big considerations. But the other one has been an attrition of Russia's global position which I think is worth tying back to. We've seen particularly the Europeans. I was looking most notably at a statement the European Commission put out with kind of a report behind it or a little more detailed informational report behind it about a week ago, making the case that a big part of this, of the current decline of Russia's operational capacity is the long tail results of policies that the Biden administration, but then now mostly the European Union have installed and maintained since 2022, primarily economic sanctions and export controls. And I think there is a kind of case to be made for this both because we've seen Russia's economy begin to collapse and this is actually the time frame that if you look talked to most experts they said Russia probably has enough stocked away reserves economic capacity that they can ride out two to four years without feeling much of a hit. But at a certain point in the multi year window it's going to get really painful for them because they will have run through their reserve, they will have exhausted their individual capacity. Particularly on the technical front. You are going to lose your pipeline of high end technology you need for various types of weaponry systems. You're going to lose the ability to develop new weapon systems, to adapt. And that is something that really hurts countries in the multi year frame. I do think there's actually a case for this. I mean this is going to be one of those big policy questions that our smart colleagues on the strategy and policy side need to be looking at saying how much did this long tail promise of these sanctions export controls deliver and how much did it not? But even the Iran case, reliance on Iran, to some extent you can attribute that to export controls because Russia has to turn to Iran because it can't turn to most other global providers for a lot of what it needs and Iran is not the ideal country you want to be buying these arms from and this equipment from, even though they are surprisingly capable on a variety of fronts, including the drone front. How much of that narrative makes sense to you, Tyler? I think I've given the friendliest possible narrative. I mean, this is what you're going to hear from European circles. You're not going to hear it from the Trump administration, but probably you will hear some Biden administration people when they don't think it will disrupt the current posture of the administration. We're afraid of that maybe as we lead up to the 2020 election, start saying, actually this policy set was successful in the medium term and long term. Does that ring true to you? Is it overstating it? And perhaps more fundamentally, are there lessons here when we look to other situations like Taiwan as to exactly how effective this toolkit is and the preconditions necessary for executing and using it effectively?
- [26:55]
E
Yeah, I mean, it's a great question and I would be curious about your answer as well, Scott. And I will also caveat. I haven't followed as closely this latest sanctions package, the Lindsey Graham named act. And so I don't know whether that'll be another catalyst or an accelerant of this dynamic. I think it makes sense to me in that the long tail of the economic sanctions package put forward in the previous administration created the conditions to allow for this, I guess, opening or movement. But I think the catalysts, to me it made more sense that the Iran dynamic, that now Iran is diminished in its capacity and its war with the US And Israel. And so I think it kind of allowed. Yeah. To create the conditions. But these other factors had to happen in order to kind of like finally break through and for the Russian economy to feel this pain. So I mean, I'm not sure what the lessons would be for Taiwan because I think it took something like the full scale invasion to prompt the US And Europe to act. I think it took this cataclysmic event rather than a. I don't see the motivation as a set of preventive measures, for example. So yeah, I don't know. I would throw it back to you, Scott. I think you'd be even better positioned to answer this.
- [28:28]
A
Well, I think what you said is basically right. I mean, we don't know this is a topic for study. The one thing I would just add to it is that that this strategy, I think it's true, probably actually did make, is making a big difference now, but it's required two things. One, sustainability, actually sustained commitment that's waffled a lot, frankly. This policy might have yielded dividends sooner if the United States hadn't waffled on it as aggressively. The Trump administration has not been as aggressive about maintaining sanctions like it has slipped a lot. The Europeans have picked up a lot of that slack. There are limits to what they can do. This new bill, we'll see exactly what it amounts to. I think there's a strong symbolic kind of threatening element. It's not clear that the administration will actually use absolutely everything in its toolkit, although it does hit an oil sector in particular, which is a vulnerability that people have been hesitant to exercise. And I think this administration will too because of the Iran war and global oil prices. But regardless, the global oil prices spiking up also helps Russia. All these factors have played in Russia's favor and we're still seeing these consequences. But you need that sustained effort. The Trump administration, while it hasn't doubled down on US and global sanctions, it also hasn't completely removed them or undermined them. And that's sort of significant. But the other thing you need is a resilient Ukraine. You need a Ukraine that's able to fight back significantly enough so that the window of opportunity that Russia was able to buy for itself, which was two to four years of relative economic stability even as it pursued a major, major military offensive, ultimately unsuccessful one, and one that would prove much more costly than anticipated, but still a major one, you got to make sure they can't capitalize on that window. And that's the hard part. I mean, that's the Taiwan question. When you think about porcupine strategy, it suggest that if the United States isn't willing to intervene militarily in Taiwan, but they are willing to slap major economic sanctions on, which is where a lot of people kind of think the United States and allies may land on it, that means you've got to get Taiwan in a position to maintain its independence for a couple of years before this kick in. Frankly, China's probably more resilient than Russia is. So that's the real question. I mean, it's got to be a two part strategy. You don't get here without the Ukrainians to state the obvious. And the Ukrainians need that support to sustain that position.
- [30:27]
D
I'm just curious. You know, Alan and I have written extensively in the AI context on this sort of policy by personality and policy by telephone and whatever the vibes are around the administration and whoever they're negotiating with and I wonder, Nastya, Ben and whoever else wants to jump in if there's a sort of fatigue of this sort of ad hoc, arbitrary policy development and negotiation strategy, there was research from the Pew Research center that came out that said that Canadians and Mexicans now view President Xi more positively than President Trump. 35% of Canadians have confidence in Xi to do the right thing, but only 20% have confidence in Trump to do that. And I'm curious if you have a hot take on if we issued a poll like that to Ukrainians. Maybe not Xi versus Trump, but just looking at Trump alone, is there sort of a growing fatigue of, hey, we ultimately are going to need to continue to expand the network of communities that are going to support us, and Trump just may not be the Reed we're able to lean against?
- [31:36]
B
Well, I don't think there's growing fatigue. I think at the moment that they yelled at our president. I mean, there was a toxic breakup and we haven't gotten back together yet.
- [31:47]
D
This sounds like a Love island episode.
- [31:50]
B
Well, it was really bad. I mean, like, I was literally in D.C. when it happened, and I was standing like near Kramer's, looking at my phone, almost crying because it felt like they were yelling at me. And I think that's how a lot of Ukrainians felt. It felt like J.D. vance was insulting you as a citizen of the country. Like, it felt like they were yelling at everyone else who lives in that country. And the reaction was really strong, like extremely strong. In Ukraine, people really felt that. And I just don't think how you walk back from it. I think ever since then, of course, the Ukrainians would welcome positive developments. And the Ukrainians are aware that the US Is still helping, giving weapons that the Europeans are paying for and giving intelligence. Everyone knows that. But there is no more fluff of, oh, we're allies. This is about protecting the democratic whatever. There's none of that. There's just hardcore interests. There is money. There's the fact that, again, the Europeans are paying for it, and that's where the relationship ends. And I mean, I guess also there are people who are waiting out until Trump is out of office and America is back to resembling what we're used to as the United States. But I think describing if we ever are, if you ever are, but I think describing it as a growing fatigue, we're way beyond that. There was growing fatigue before the Oval Office meeting. I think at the Oval Office meeting, it just flew flipped.
- [33:14]
A
So let me go to one last part of this for you, Nastya, that I think is really important as we think about the next step, where we go from here. We had these reports that there is a new ceasefire proposal in the offing, beginning of a new kind of on ramp to some sort of peace negotiations. I think the ceasefire is a more realistic medium term goal as opposed to a full resolution of the conflict, but something to ramp down the scale of conflict. Sounds like Zelenskyy is on board. He's working with Trump administration to develop these sets of ideas. And we know President Trump has said, you know, completely unrealistically since before he was reelected. Oh, yeah, I'll come in and I'll settle the Ukraine conflict in a month. Right. Obviously that hasn't happened. Obviously that was ridiculous thing to say.
- [33:55]
C
24 hours, not a month, was it 24 hours.
- [33:58]
A
I couldn't remember exactly what unrealistic timeframe it was.
- [34:00]
C
He said it many, many times. He now contends he was joking. Go watch the video. He was not joking.
- [34:08]
A
It is absolutely wild, the expectations, but the interest still seems to be there. I mean, this is President Trump that in spite of Iran, still describes himself as a peacemaker and at a more fundamental level I think actually does say, I don't like this ongoing tension with Russia, this ongoing conflict, finding a way to take this off our table and in the worldview of him and certain people in his administration, take away this friction point with one of the other two major powers that we need to come sort of accommodation with and kind of splitting the global system. Right. If you follow their national security strategy sort of logic. The question I have for you is how much is this earnest on the part of Ukrainians and how much is this playing into the playbook that they think Trump will accept? Because we know during the last minute of real optimism for the state of the conflict in 2023 that led to the spring and summer offensive, which was I think still viewed as not a success and a mistake, but there reflected a very real drive to say we want to push back and make more gains on the ground against Russia, and there's still plenty of reason to desire that. Now, is the Ukrainian an appetite for a at least temporary settlement, a ceasefire on the table genuinely, or is this reflecting Trump's preferences about where he wants things to go? And Zelensky and other people around him are just saying this is the menu opportunities. If we're going to keep the Trump administration in our corner, we've got to play in the direction that they want to move in.
- [35:31]
B
So it's actually a complicated question. I think it's probably more complicated than many people realize. So I think there are many sides to this. So first of all, as you've said, Trump has been pushing for a ceasefire just so he has something to sell to his domestic audience. And therefore the Ukrainians picked up on that and said, yes, we're pro ceasefire, ceasefire now, unconditional, we're gonna agree to anything. And I think Ukraine has been kind of still on that side and saying they're ready, they're ready. So that's one part of it. The second part of it is, of course, that the Russians are vehemently against any sort of pause in fighting. And so the Ukrainians continue saying that they're pro ceasefire, which then ultimately highlights the fact that the Russians are the problem because they don't want one. At the same time, I think it's the question of whether Ukrainians genuinely want a temporary pause of fighting is very interesting because there is a lot of evidence to show that A, the Russians don't actually abide by the ceasefire. But once you've entered it, it, once Ukraine and Russia enter into it, you've got to abide by it because optics matter. And you're again in that optics game of like, okay, who is Trump gonna blame? So if the Russians keep breaking the ceasefire, we kind of have to continue being the good guys and we have to continuing to abide by it. And that's awkward and not optimal on the front line. And then also there is plenty of evidence that the Russians use the ceasefires to gain tactical advantage to regroup. There have been papers on it. And there is generally all sides in
- [37:07]
C
all conflicts do that.
- [37:09]
B
Okay, sure. All sides in all conflicts do that. Yes. And so you could also argue that the Ukrainians are not really interested in a ceasefire, in a temporary ceasefire. Like we're going to take a month off and then we'll continue this because what's the point? This is only going to give Russia more time to regroup. And since Ukraine now has the initiative, since Ukraine is now kind of doing really well, there was really no point, I guess one could say, in stopping that. You could then of course also argue that yes, the Ukrainian long range strikes are doing really well and the medium strikes are doing really well. But let's not forget about the Ukrainian infantry who've been sitting in holes in the ground for many, many months in horrific conditions. And then they haven't been able to rotate them. When I say that Ukraine is doing well, I'm not dismissing the fact that the front line is still a horrible place base and the Ukrainian forces are immensely stretched, extremely thin. And so it's just a complicated question and there hasn't been a lot of public communication. I mean, this is one of the biggest problems with the Ukrainian government in general. The Ukrainian government doesn't talk to the Ukrainian people. The Ukrainian government talks to the American media, to Laura Loomer, to the American Congress, to the Europeans. It does not really talk honestly openly with the Ukrainian society. And so I'm not sure if the Ukrainian top generals and the Ukrainian government genuinely want to cease fire or not. There are a lot of factors here. What I do know is that the Russians are sure as hell not going to agree to any of it.
- [38:45]
A
Ben, what do you make of this? In some ways, I think when we talked about this podcast previously, we said, strategically, Zelenskyy needs to look to Trump to keep Trump in his corner, like the least problematic of the two actors. And early on, Putin was playing the game well by responding to calls, taking calls from Trump and saying, of course I'm open to a ceasefire, absolutely happy to talk about it, let's get direct negotiations going. But those lines of communication don't seem to be bearing the fruit that they once did for Putin. And frankly, even his public posturing has become less acceptable to those sorts of things. And the Trump administration has responded in kind. What is the genuine endgame here? I mean, this was always a war of attrition. It was which side was going to be outlast the other and make it more painful for the other. Has the balance shifted enough that it now weighs in favor of Ukraine, at least for a window here? Or is there a reason to think that there's a driver towards a resolution of this in the short to medium term that makes sense from the Ukrainian perspective.
- [39:41]
C
Let me be really blunt about this. There is no prospect of a ceasefire under current conditions, none whatsoever. And there's two reasons for this. The first is that Russia doesn't want one, and the second is that Ukraine doesn't want one and you need both of them to want one in order for there to be. Now, why doesn't Russia want one? Because Russia is operating under the sunk cost fallacy. You've thrown a million people and gotten literally a million people killed and you have absolutely nothing to show for it.
- [40:21]
B
It's a million killed and wounded.
- [40:23]
C
It can't kill. Yeah, a million casualties and you have zero to show for it. You have Moscow burning, you have St. Petersburg burning. You can't stop now. Now Ukraine, as Nastya points out, has a real two faced posture here. Their public posture is kind of like Eddie Felson, the Paul Newman character in the Hustler. Right. The game is over when Minnesota Fat says it's over. Anytime he wants a ceasefire, ceasefire in place. Sure. They're saying that because they know that Putin can't agree to it and because it is a great public posture for them to have. Have dirty little secret. The side that perceives itself as progressing never agrees to a ceasefire. Never. If you're gaining ground, if you're making the other side sweat, if you think your posture is going to be better in three months than it is today, you don't agree to a ceasefire. So the sort of Fast Eddie Felson, the game is over when Vladimir Putin says it's over is a. You know, it's a ruse.
- [41:39]
B
Which also means that the ceasefire. Sorry, I'll interrupt you. The ceasefire now is actually a lot less likely than it was a year ago.
- [41:46]
C
A year ago you could say it was going to happen because the Ukrainians were going to have to agree to it under very undesirable terms.
- [41:54]
B
Exactly.
- [41:55]
C
Now it is not. That's just not going to happen. And so there is only one circumstance in which there is a plausible ceasefire, and that is the circumstance that ends the Iran Iraq war. If you go back to, I want to say, 1988 or which is, you know, you have this long stalemate kind of thing, and then all of a sudden the Iraqis break through and the. And Ayatollah Khomeini is forced to. To ask for a ceasefire and goes on television and says, this is the bitterest moment of my life. I've had to ask for a ceasefire. In other words, one side has to lose or has to be in imminent danger of losing and has to be forced to ask for a ceasefire. That is how this war is going to end. It is not going to end because they kind of tire each other out, that may lead to a major decrease in the pace and rapidity and intensity of the conflict. It may lead to something not frozen, but defrosted a little bit. But a ceasefire, Somebody's gonna lose this fight and that's when it's gonna end.
- [43:21]
A
I don't fundamentally disagree with that. I agree with it. But what I will say, though, is that last point, that caveat about the lower the tempo. I think a significant one here, because I think the dynamic of this conflict may have shifted enough where the timeframe benefits the Ukrainians. Meaning if you can make it less painful for you in the interim and let Russia's own internal inconsistencies catch up with it more that's beneficial. That was not true earlier in this conflict, before we knew Russia was able to was going to feel this pain. It may have been true was all hypothetical because we saw the the policy consequences of a lot of sanctions, things like that were years away in terms of the ability to recover from a lower tempo of hostilities. I'm not sure if the balance hasn't shifted to the Ukrainians, where I think the presumption has been that the Russians would benefit from that more through most of this war. If Russia's economy and industrial capacity has taken the hit, it seems like it has and domestic pressure is building enough. So that's the one caveat I put on that that may not be a ceasefire, but it may be reason to say we can tamp down to scale this conflict because it's more about outlasting than besting at this point for the Ukrainians and they have a reasonable window to outlast. Does that sound wrong to you?
- [44:28]
C
Yeah, no, I think that sounds exactly right, but with one really important caveat to your caveat, which is that you can decrease the tempo at the front. But as Nastya pointed out before, there are tools other than the front. So the Russians have bombarding Nastya at night, right? And that is their go to. Disrupting Ukrainian civilian life is their go to. For the Ukrainians, the go to is long range strikes that really substantially affect the Russian economy and also their ability to feed the front. So one perverse possible consequence of lowering the tempo at the front is raising the tempo elsewhere in both countries.
- [45:20]
A
Hey folks. Scott R. Andersen here. It is officially hot here in Washington, D.C. the kind of hot where biking to work means arriving already regretting your life choices, where the only sane afternoon plans involve shade, something cold to drink, and anything other than jeans. Luckily, Quince has you covered for the season. I mean that quite literally, because half the trick to surviving a D.C. summer is having the right lightweight stuff in rotation in your wardrobe. Personally, I have been living in their organic cotton Coolmax Juno shorts this month. Breathable, comfortable, and somehow still presentable when I'm actually biking and running around. Their ultimate commuter shorts have become my go to. They move with you and don't fall apart after a few washes. And their Pro Peak Performance pillow has quietly become my favorite warm weather shirt. Looks sharp but breathes enough that I'm not melting into my chair. And when the weekend rolls around, it's pool time with the kids. I'll admit I've gotten a lot of mileage out of their Italian swim trunks genuinely nicer fabric and fit than I expected for swimwear. So good, in fact, that I just bought another pair. The best part? Quints works directly with ethical factories, cutting out the middleman markup so you get real quality without the premium price tag. So beat the heat this summer with Quince. Go to Quince.com Security for free shipping on your order and 365 day returns. Now available in Canada too. That's Q-U-I-N-C-E.com Security free shipping and 365 day returns. Quince.com Security now let's get back to the show. Summer smells like salt in the air and warm sand. Restore your sense of place with Pura's new summer fragrance collection. Discover transportive clean scents@pura.com the snack wrap
- [47:09]
C
plus Caesar sauce equals extra crispy chicken Caesary Greatness.
- [47:14]
A
Sounds delicious.
- [47:16]
C
Caesar sauce at McDonald's for a limited time.
- [47:24]
A
It's a fascinating moment in this conflict that we follow so closely here on other channels, and we're so lucky to have you there, Nastya, talking to you about it both from your personal experience and doing journalism on the ground. We'll have to move on to another topic now. In fact, we may have to drop a topic given how long this conversation has gone on, but I'm okay. I think it's a really good conversation. But let us move our attention now to another situation that has gotten thoroughly out of control in recent weeks in a way that some people seem, including the affected people, kind of delighted by in an interesting way, or excited by where other people are absolutely horrified. Kevin, I want to come to you on this. This is of course the hugging face incident, which at this point I think probably most of our listeners have heard about but may not fully understand the full details of what happened, despite my efforts with the sentence I read at the introduction. Can you give us a little sense about what is exactly that happened here to clarify it for people and talk about significance? Why is this such a big deal that this took place, that it has so many people responding with howls of elation and howls of terror across the community? People who watch AI Law Policy?
- [48:34]
D
Well, first and foremost, it's important to note that no one should know exactly what happened because OpenAI hasn't disclosed exactly what happened and shared all of the information about its logs, how it was running this test and what transpired as a result. So hopefully we will have that information soon. Breaking news as of last night is that Meter and Redwood Research, two independent nonprofits who focus on AI analysis and AI evaluation are going to do an independent audit of this incident and share some insights into what actually happened behind the scenes. So hopefully we'll be, we'll get some more transparency soon. But to just go back to what the hell happened, what's going on here? For folks who have lived under a rock or in a nice paradise in which AI drama doesn't invade and dominate their news feeds, there's a process that labs go through internally that's somewhat akin to crash testing a car. In the same way that we often see car manufacturers will take their car, put it in a safe setting, you know, some far away destination, maybe remove some of the traditional standards and safety mechanisms they would have in place and then drive that car at full speed into a wall and see what happens. That's somewhat what we see from AI labs themselves. They will take their models, including models that have not been deployed yet, remove some of the safeguards that they would otherwise impose, and put them into sandboxes and say, let's see just how capable this model is on some of those risky behaviors. So in this case, OpenAI was testing the cyber capabilities of one of its models that it hasn't released yet and was engaging in its ability to complete what's known as the exploit gym. This is just a test of its cyber capabilities. And this was supposed to be a test that, like a car crash test was closed and not something that was going to leak out into the open streets or onto the open road. And yet what we saw happen was this model was able to find various mechanisms to escape that sandbox. This is a loss of control scenario in which suddenly the lab expects that the model is going to operate in one way and perform a certain way within its safeguards. But it escaped those bounds. And it escaped its bounds in a way that OpenAI didn't immediately realize, which led to a hack of Hugging Face. So Hugging Face is a repository of OpenAI systems and they suddenly reported to law enforcement that we've noticed in highly capable agentic AI system is hacking into our backend. They reported this to the authorities and said, we're not sure who's doing this, we're not sure how it's happening. And eventually it became clear that it was indeed OpenAI that had perpetrated this cyber intrusion into Hugging Face. The model was able to get into Hugging Face's systems, exploit some zero day vulnerabilities, move laterally throughout the Hugging Face interface, and fortunately did not cause, as far as we can tell, any massive damage of financial concern or of informational integrity. But critically, Hugging Face, in response to this incident, tried to go and use the latest and greatest AI tools, including Anthropic's latest tools, and say, hey Claude, help me respond to this highly sophisticated AI attack. To which Claude said, you know, go fish. No thanks. This looks like the sort of science cyber queries that we don't want to necessarily sustain and allow. And so instead Hugging Face had to turn to open source models, namely GLM 5.2 from China to respond to this attack from OpenAI. And so the result was OpenAI leading to some degree of publication of how and why this transpired. There are a lot of questions and a lot of missing gaps in what actually transpired from OpenAI side for example, why this wasn't being more closely monitored by their employees, how their sandbox could have been better safeguarded, why they weren't using, for example, an air gap system, which tends to be more expensive and may not allow for the full degree of capabilities testing that OpenAI would would want to see. And also getting a better sense of what does this mean for broader debates in terms of who should have access to the latest and greatest AI and at what time. We now know that Hugging Face has been added to OpenAI's list of trusted partners who will receive access to their models sooner rather than later to make sure they can take these defensive steps and secure and harden their systems critically. Also, in addition to the news that's broken about Meter and Redwood research doing a independent analysis of this incident, it's worth noting that Sam Altman flagged that he's not sure if other systems may have been hacked besides Hugging Face as a result of this breach of the sandbox. And we've also seen research from the good folks at the Information, namely Rocket Drew, noted that we've seen that this model actually left notes for future models about how to escape the sandbox. And so all of these disclosures and all of this information, to your point earlier, Scott, has caused quite the just mixed reaction across the AI community. On the one hand, we've seen some groups, even Hugging Face's own CEO, say, whoa, this is really cool. Look at how capable these models are. This is proof that that we are creating highly agentic systems, systems that are very well equipped to pursue goals and achieve highly sophisticated plans to achieve those goals. That's wild from a technological standpoint. If you had told anyone about this incident even three years ago, they would have said, whoa, you've got some aggressive timelines there. On the other end we've seen, thanks to excellent analysis by, for example, our own Kate Clonic. This should be wildly concerning to anyone who is not a part of a trusted partners program, right? If you're not getting access to the latest tools to take those defensive measures sooner, well, gee, all of a sudden you're left asking, just how sophisticated is the sandbox that OpenAI or Anthropic or Google is using to make sure that their internal testing doesn't allow for this sort of hacking down the road? Now, hopefully, again, as Kate and others have pointed out, this should be a rallying cry for we need way more transparency. Right? It shouldn't be a guessing game about what information is going to be disclosed and when. We need way more standardization, such that when you are transparent, we know what information you're going to share with the public, such that there can be a more nuanced response. And number three, something that I don't think has received enough attention is AI testing is really hard and really expensive. And so if we want better, safer, more reliable AI tools, we shouldn't just expect that the labs themselves are going to be capable of doing this or necessarily have the market incentives to do this. This is where Congress needs to buck up Its like 35 days remaining and pass meaningful legislation about making sure we can fund and support a meaningful eval environment. But those are three quick takes. There's a lot to dig into and, and this is certainly going to be a story that earns another chapter in the history of AI.
- [56:28]
E
Kevin, that was very comprehensive and very helpful, but one question that I have that you didn't answer that we're glossing over is the fact that this organization is called Hugging Face. Why is it called Hugging Face?
- [56:40]
D
You know, that's a great question. I'm not sure. I'll have to ask a former classmate of mine, Irene. I know. Is that hugging face? I will get back to you. Tyler and I will do better research next time.
- [56:53]
E
We keep throwing around that name and no one is really acknowledging how strange it is.
- [56:56]
A
I think it's off of the Bitmoji or not Bitmoji, the emoji.
- [56:59]
D
Right.
- [57:00]
A
That's kind of like they've got a great emoji.
- [57:02]
D
Yeah, yeah. Their emoji is quite literally a hugging face. I don't know why they picked a hugging face vibes. Who knows?
- [57:09]
E
Well, I, I love that. I. The only thing I would add is, is another shout out to Kate Clonick's great piece in Lawfare from, I believe it was July 29th to Kevin's point, she represents the argument of if the model broke out of the sandbox, then maybe the sandbox is not so good argument. But I thought another really interesting strain of the piece was essentially making the argument that another complicating factor here is that the tech companies themselves aren't the most reliable narrators or there's a lot of good reason to question in the narrative that they're presenting because the idea that OpenAI orchestrated this as some grand marketing scheme seems very far fetched. But it does make sense that they don't want to waste a good crisis and she drilled down on what little disclosures they have made or essentially their statement, which kind of is this same playbook you see where some sort of crisis happens and and the companies are just sort of in awe of what they've created, which is pretty good marketing. Wow. The very people who are making it are a little scared of what they're making. It must be powerful. You see this again and again. Again, not to say this was some brilliant 3D chess marketing guerilla marketing campaign, but it does make sense that they would use it to. It's a lot easier to say that wow, it's so powerful what we built than it is to say, oops, we left the keys in the door. You know, it's from a PR perspective,
- [58:47]
D
I think there's certainly a lot of interesting discussion going on around what are the PR benefits of this approach and was this even something that was coordinated with Hugging Face? Now, a lot of smart folks have pointed out that, you know, Hugging Face is a champion of open source. It would be weird for Hugging Face to necessarily be celebrating this specific incident where they tried to use a closed model to defend themselves and it didn't work out as intended. I do think, though, this needs to be framed in prior examples of tech policy where we've seen horrible incidents occur as a result of a company's perhaps lax approach to safety. And the response has been what people refer to as transparency theater. And my biggest fear is that the response to this is going to be just share more information, come up with more documents, post more things on your website, and we're not going to see the sort of substantive investment in AI evals and AI testing that we really need to see in response to this because we already saw, as folks have pointed out in OpenAI's own statements, they said that, hey, when we see this kind of incident occur, we're going to pause development, development. And whether you support that or whether you want to see that is a separate question in their own documentation about the sorts of steps they were going to take in response to new developments to AI they had, in theory, set themselves up to now have a real pause moment. Now, I think Sam has specified that they are reconsidering and making sure that their testing environment is strengthened and hardened. They've outlined a couple of steps. For example, increased logs that they're going to monitor with respect. Respect to their sandbox, increased mechanisms for ensuring that there are reports when that sandbox may have been breached. And that may have been the sort of immediate pause that was taken. But the broader discourse has to be about creating an ecosystem in which this isn't just so based on ad hoc decisions by a lab to disclose information and then pick two random nonprofits out of the blue and say, these are the two we trust now to do our independent evaluators evaluation. We need to be much more structured and deliberate about this.
- [61:09]
C
I want to just point out a point that is latent in a lot of what Kevin has said, which is when he talks about, isn't it cool? Even Hugging Face acknowledges that we have these highly adjacent systems that can do cool things like attack us. What he's really talking about is nascent. In those words, highly agentic systems is the idea of an AI with will. And we think of AI as a thing that, before you prompt it, is just entirely dormant and then springs to action to do what you want, or to answer your question, or to do the coding that you've instructed, or to make reservations for you to get on a plane and then get into a hotel. But, you know, in order to do those things, it has to make all it has to want to. Right. And the reason it wants to is that you tell it to want to, and then, you know, at what altitude it wants things and can make decisions to break into. You know, oh, I can't make a hotel reservation for Scott unless I break into the hotel hotel's system. Right.
- [62:31]
E
Scott did ask it to do that,
- [62:32]
C
though, you know, and. And you didn't ask it to hack the hotel, you just asked it to make a reservation. But it had a problem when it made a reserve, tried to make a reservation, which is that the hotel is booked. And so what it actually needs to do in order to effectuate your will and it's not consulting you about this, is to hack the hotel system, cancel the reservations of some other family, and then make your reservation. That's a rough analog to the hugging face attack. Right. But then, of course, the Question is, at what altitude does it have will? Right? Does it merely have will to effectuate Scott's desire? And is that at a sufficient level of altitude that it might do dangerous stuff, or does it at some point get to say, actually, Scott's really asking the wrong question here? Here he shouldn't want to be going to a hotel, he should be wanting to stay in an Airbnb, you know, and make, you know, supplant human decisions with its own decisions. And I think part of what is really dangerous about this is that we have, we have no agreement as a society about at what level AIs should have will. Should they merely have will for purposes of effectuating human judgments and desires that are that it can do without endangering anybody. Should it have will at a high level of agentic autonomy? You know, go make the world better, Claude.
- [64:11]
A
Right.
- [64:12]
C
And then it decides that certain countries are really a problem for making the world better, so you nuke them. Right. Or should it have really autonomous decision making will? And those raise very, very different security environments. And right now, other than the assumption that there is a tort system that can deal with some damages, we have no agreed upon sort of social answer to this question. So I think that I raise that not because it's immediately raised by the hugging face face incident, but because I think it's the sort of tectonic layer of the hugging face incident.
- [64:58]
A
I think that's really useful, Ben. And what's really interesting about this incident, Kevin, I want you to help me correct me or fill my gaps. Am I missing this? The part that's really interesting to me is that it's essentially a tale of two different sets of constraints. External constraints at the sandbox which failed. And that again could just be OpenAI not having done a good enough job building a sandbox. Right. Presumably they get to use the same model to look for zero day issues in the sandbox before they ran this test. And they could have eliminated it and preserved it and therefore they wouldn't have escaped. And maybe that's a good thing to do moving forward. Synthetic effect, Right. If you don't want to do the air gap system, but then you have the internal constraints which were deliberately turned off for this model, as I understand the description I've gotten, is that essentially they said the internal constraints that would normally stop these systems from doing this, the built in ethos, ethics, morality, but not really, that's like the human parallel that would normally exercise self constraint for the model, we deliberately turned that off. And then those Same things became a problem when they tried to use Fable to fix it because that's where the internal constraints anthropic is built into. Fable said, whoa, whoa, whoa, whoa, this is too close to weird cyber stuff we're not supposed to be doing. We're not going to go into it. That's why they had to go to GLM 5, whatever, 5.2, I think. Right. So the thing that I think is interesting for me, this is a, that whatever this phantom model is or the ChatGPT model is, it's the fact that they built that they roadmapped the way out of the sail sandbox. That's the part to me that's so interesting because it's not directly pursuant to the immediate task that it's executing. Right. Like presumably you said, we're giving you a test answer this test. They go to hugging face, they break it, but instead they left this trail of breadcrumbs to let another system out in the future. Now that actually makes sense if you interact with these AI systems because they're always learning from what you do. Every time you ask it to do something, it builds these sorts of trails to make it more effective the next time you ask it to do something. And that's actually a great thing. That's part of the reason they're so effective is because the more you use it to do something, the better it gets at it learns. But it underscores actually kind of the weird consequence of that because you're saying, okay, but when you make a mistake, you kind of go beyond it. But more fundamentally, I wonder whether this really underscores the importance of that internal. I know anthropic calls. There's the constitution, the internal guidelines, the standards, not rules that it has to run up against because these systems are going to be so effective at breaking the rules and moving past them. And then I think the real question then becomes, well, how do you build in a set of guidelines of internal restrictions that as brittle as external rules Any law student has encountered the way you can reason and logic around internal rules if you're really pressed on it. Ben's had fun time doing that with early language models. I remember when he got it to say, I think chatgpt say some horribly anti semitic things it was supposed to not be able to say, but if you provoked it and asked it questions the right way and worked around some hypotheticals, you can massage it into doing it. How do you develop a system that combines recursive logic, sense of overall impact, moral fencing, all these things that we use to kind of rein in those arguments in the real world to keep those internal guidelines on track or maybe to start alerting people if people start pressing against them. And what's crazy to me is that's actually. While GLM5.2 comes out as the hero in this because they solved the thing, isn't that what makes those open source models so scary is because those are exactly the sorts of models where you can turn off those internal guidelines a lot more easily than these closed anthropic chatgpt models. That to me actually makes them way. The takeaway from this story, the more I thought about is that, man, I'm terrified of those other models. If the thing that really seems to have kept these models in track, except in this one instance, is these internal guidelines that makes these ones that you can turn those things off way scarier or that anyone can turn those off, not just the company providing it. Am I wrong about that, Kevin, or do you come away with a different track on this? That, to me is actually where I came along is much more scared of these open source models, which I've been playing with and I'm intrigued by and I think excited by on a lot of, of like economic diffusion model, like a lot of applications, that actually makes a lot more sense. But putting that sort of capability in a way that anyone can turn off the guardrails. I don't know. This incident scared me from that perspective.
- [68:57]
D
So I'll try to keep my remarks to 70 minutes in response to that great content. First, I just want to. On Ben's point, I don't think that in D.C. and in a lot of state capitals, I'm not sure that folks realize what it means when a, a AI employee says something like I'm going to spin up 100 agents to do this task. And I think if you asked a lawmaker just off of that sentence alone, what is this employee saying? They'd have no idea.
- [69:28]
C
Yeah. And the word agentic actually is obfuscatory in this respect. It's like you're talking about art. We don't use the phrase because it sounds super creepy creepy. But what we're talking about here, we love the phrase artificial intelligence. We're talking about artificial will. And once you say a computer with will, people understand that you're talking about something scary. Yeah.
- [69:56]
D
And this is where I think we need such a more deliberate effort to build off of efforts like the Frontier Model Forum, like other independent bodies that are bringing together lawmakers and the technical community to explain what the heck is going on because until we have that level of sophistication among our regulators, we're not going to get the sort of nimble, thoughtful, evidence based regulation that could be the scaffolding for this. And Scott, this is relevant to what you said in the conversation about how to develop better testing. The reason why we saw this model go to Hugging Face was because it's been trained on data that suggested, hey, hugging face may have these answer. And so if you want better testing, you need better data that the model hasn't seen before, for example. But that's expensive, that's hard to do, that's a market failure that we really need to have somebody that's investing in that kind of data and allowing for more thoughtful and clear and for lack of a better phrase, clean testing such that there, there isn't this impulse to cheat by going to hugging face, for example, and trying to find the answer. Now, on the open source question, this is where I think we really could have a 70 minute long debate. You know, absent open source right now, who knows what could have happened to hugging Face, what the long term ramifications could be. Or fill in the blank for any other target that had been the hacking target of OpenAI's model. In this case, open source was the thing that allowed for that cybersecurity response from hugging face. Now the question is how to make sure that we have a pro defensive leaning use of those open source models as opposed to a offensive capability use of those open source models. Now, all else equal, we've seen that the history of technology and of open source generally has been that open source is incredibly helpful in allowing for more researchers to do more work into vulnerabilities and safety testing and research in a way that isn't feasible right now. So as an academic at a well resourced university, I'm blessed because a lot of my colleagues in the CS department at UT can go and get access to the latest and greatest closed systems. But for the rest of the academic community and the independent community to do that research, Open source models are really helpful there. So I don't think that a rushed response of banning open source, for example, is where we should be headed. Anthropic, for example, released a statement trying to specify that there needs to be a thoughtful middle ground of not banning open source while acknowledging that there are very real risks like you mentioned, Scott. And so this is where having an adult conversation about AI has to happen. And that's just not taking place on the hill. Right now we're debating you know, 290 page pieces of legislation rather than, in my opinion, what should be treated as like a policy hackathon down. Let's just address discrete, specific topics and move on and act like lawmakers should in this domain, which is to say, make actual progress, decide on standardized disclosure forms, identify independent auditors, and really start to move the ball towards a more robust testing environment.
- [73:16]
A
Well, it is a fascinating topic, one we could talk about a lot longer, but we are out of time today. In fact, we didn't even get to our third topic. But that's okay. Sometimes that happens. We'll save it for next week, but this would not be Rational Security if we did not leave you with some object lessons to ponder over in the week to come. Tyler, what did you bring for us today?
- [73:34]
E
I brought a. A documentary from 1986. The full name is Sherman's March Colon, a meditation on the possibility of romantic love in the south during an era of nuclear weapons proliferation.
- [73:46]
A
I literally. I just started watching this the other week. I didn't get to finish it.
- [73:49]
C
I love this movie.
- [73:51]
E
I was gonna particularly. I was curious if Ben has seen it, and I'm glad you have. I'm bringing it up because 4K restoration was just released and they've been doing some screenings at Film Forum in New York City. Unfortunately, I was not quick enough on the draw to go to one of The Q&As that was moderated, the first of which was moderated by Ira Glass with the filmmaker Ross McIlwee. And the second was Zoran Mamdani's mother, who's a famous, very famous filmmaker in her own right. And I wish I could have seen either one of them. But the 4K restoration is amazing. For anyone who hasn't seen it or know about it, the quick, very quick setup. And all I'll say is this quick setup is that essentially this documentarian from the south got a Grant in the 80s to document the lasting effects of Sherman's march to the sea through the South. But shortly before he departs on his trip, his girlfriend breaks up with him. So he just kind of unravels and he becomes enamored with different Southern women. And it's an exploration of those relationships.
- [74:51]
C
And everywhere he goes to film Sherman's March, people set him up with Southern women. And so it's. It's a no. It's fabulous.
- [75:00]
E
It's incredible.
- [75:01]
A
Well, and the reason why I think the release came out is because he. I think it was last year, technically released a new film remake, which is. Pulls together A bunch of footage from his child about the death of his son that looks, it's supposed to be really, really striking. I mean, his son, I think wrestled with a variety of kind of substance abuse, mental health issues, but he kind of traces his childhood through these incredible videos. It sounds amazing. I have not mustered the willpower to watch that movie. But I started watching the Sherman's Movie, Sherman Movie, Sherman's March because I was like, well, that sounds up more upbeat. That one I couldn't bear with. I think I'm going to have to steel myself to watch remake. But it's also supposed to be amazing. I've read a couple of reviews saying it's one of the best movies I come out the last few years.
- [75:41]
E
Years.
- [75:41]
A
Even though it's gotten really overlooked, it seems like even in festivals and stuff. So I'll throw that in as well as Ross McIlwee I think is how you say it. I've never said it out loud, but is the filmmaker phenomenal? Recommendation. Anastia, what do you have for us this week?
- [75:54]
B
My recommendation is a little basic, but I'm gonna recommend George Orwell's Homage to Catalonia. And it's very much related to what we discussed. Obviously it's one of the most famous books about war and what it's like to be fighting a war and living amid a war. But I remember reading it recently and think but it gave me this very weird feeling of validation that like, okay, all of this crap that Ukraine is going through and all of the various shortages and the government system not working and the bureaucracy not working well enough to address this horrific thing that is a full scale war, that every country at war goes through that. Which is maybe a very basic thought, but I found it very therapeutic to read about the conditions that people were living in and the military and all of those problems. And I was, wow, it really is a kind of a universal war experience. And maybe I should tone down my 247 criticism of my government. I didn't though. I didn't. I thought I did not. But great book.
- [76:57]
A
We can say from the confidence that's not the right way. Stick with it. Ben, let me turn to you next. What do you have for us this week?
- [77:05]
C
So I have a data set that I have not yet created, but it's going to happen this afternoon. And this is based on using the tool Ragtime. I decided to set out and solve one of the great unsolved problems of American law, which is to count the number of federal crimes. And this problem has been lurking around the American legal system since the early 90s, 1980s, when the Justice Department tried to count the federal crimes and failed, and the ABA then tried it and couldn't count them. There have been a number. The Heritage foundation teamed up with the NACDL once to try to count them and they failed. And I think that ragtime can do it. And so today we are doing phase one one, which is not the number of federal crimes, it's the number of statutes that create federal crimes. And by the end of the day I am going to have a number and I'm going to publish it on lawfare next week. That reflects part one of what our colleague Katherine Pompilio has dubbed rag crime, which is the application of ragtime to the problem of counting federal crimes. By the time you hear the next Rational Security episode, you will be able to look up the number of federal criminal laws in the United States.
- [78:43]
A
You know, I have been listening to all these, reading all these profiles of winners of the Fields Medal, so this doesn't seem like that impressive math problem, but for a lawyer or legal folks, I guess it's not that bad. So we'll take it. So that's good. I'm excited to see the results. Kevin, what did you bring for us this week?
- [78:58]
D
I've got a heartwarming story of a bunch of students getting together thanks to the Edward Kennedy center and doing a deliberation on how AI should be governed in their schools. And so they got together thanks to Day of AI, which is a part of MIT, and deliberated and crafted their own rules for how AI should deployed in their school system. So a great example of folks getting together around a targeted question, coming to a consensus and issuing some idea of how to move forward and seeing the next generation take that on is really exciting.
- [79:36]
A
Wonderful. What another another great object lesson. Well, for my object lesson to bring it all home for everyone, I'm going to bring a pair of object lessons, one from the absurd, one for the more serious. Like so many. I have been playing with AI a lot lately, working a number of research projects. I'm helping Ben with various rig time related initiatives and efforts and it's just an amazing tool but also does raise these questions about what is our added value in a world where these tools are only going to get better in the near future. And my thinking keeps coming back to a piece of writing that I revisited that I frankly refine myself revisiting like every year or two. I think it's fair to say and I think I have said, said maybe on this podcast that it is one of the three or five most influential things I've ever read on the Way, I think, and I do find myself referring to it time and time again. It's particularly relevant to this conversation, which is Max Weber's Science as Vocation, the far more overlooked, far less discussed companion lecture. It was originally a lecture that he delivered to Politics as a Vocation. It is an amazing reflection on the role of rationality and the impact of rationality on the human condition and humanity. Experience and AI is in the end kind of like a mass implementation of rationality, perhaps even independent of the human experience, at least a form of rationality. But it's interesting where he comes out, which is not necessarily like a super hopeful vision. Remember he was writing in kind of pre war with Germany or kind of in the midst and after World War I. But it has interesting lessons, I think, for humans as we think about how we can add to and bring out a value even in the age of AI, which I find kind of of enlightening about the limits of rationality and what it can and can't do for human beings and where we turn when to answer other more fundamental questions that rationality can't reach. I wouldn't say it's a happy story necessarily, but it's an interesting one and very provocative and one that I feel like is a good length through which to view these sort of contemporary questions. The other object lesson I will pair is the best SNL skit I've seen in a long while and that is Ariel Tramway v. Red Heart. I think it's called a bit from Weekend Update. That's all I'm to going to say. You have to check it out. Mikey Day, though is definitely a comic genius and is one of my favorite things I've watched online in a long time. So go ahead and Google and check that out, particularly in your moments that you may be having a little bit of existential dread. Well, that brings us to the end of this week's episode. Rational Security is of course a production of Lawfare, so be sure to visit lawfaremedia.org for our show page for links to past episodes, for written work and the written work of other Law Firm contributors, and for information on Law Firm's other phenomenal podcast series. While you're at it, be sure to follow Lawfair on social media wherever you socialize your media. Be sure to leave a rating or review wherever you might be listening and sign up to become a material supporter of Lawfare on Patreon for an ad free version of this podcast, among other special benefits. For more information, visit lawfaremedia.org support our audio engineer and producer this week was me of me and our music, as always, was performed by Sophia Yan. We were once again edited by the wonderful Jen Pacha. Behalf of my my guest Ben, Tyler, Kevin and Nastya. I am Scott R. Andersen and we will talk to you next week. Till then, goodbye.
- [82:50]
D
Most people don't realize how much of their personal information is being bought and sold every day. Data brokers are making billions, pulling details about you from public records in the Internet Internet, then packaging and selling it, usually without your consent. That's how your information lands in the hands of scammers, spammers, even stalkers. It's why you get endless robocalls and
- [83:11]
A
why ads seem to follow you everywhere.
- [83:14]
D
That's where Aura comes in. Aura actively removes your data from broker sites and keeps it off. They also instantly alert you if your information shows up in a breach or on the dark web. But Aura goes beyond data protection. With one app you get a vpn, antivirus, password manager, spam call protection, dark web monitoring, and even up to $5 million in identity theft insurance, all backed by 24. 7 US based fraud support. Other companies might sell just credit monitoring or just a vpn.
- [83:44]
A
Aura gives you all of it together
- [83:46]
D
at the same price competitors charge for just one service. Start your free trial today@aura.com safety protect yourself now@aura.com safet.