METR Investigates OpenAI-Hugging Face Incident and California Legislature Advances IVO Framework

[music] Welcome back to the AI policy podcast. I'm Oluka, director of the Wadwani AI Center here at the Center for Strategic and International Studies. >> And I'm Nicole Herrera, a researcher with the center. Over the past two weeks, the Open AI hugging face cyber incident has continued to dominate the AI policy news cycle. Both OpenAI and Meter published findings from their respective investigations, which we're going to discuss in depth today. But there are a few stories I want to make sure we cover first. The California State Legislature passed an IVO bill this week. Bill Gates published his take on the current moment in AI, and a federal judge struck down the Trump administration's move to label anthropic a supply chain risk. We've got a lot to cover, so let's dive right in. On August 30th, the California State Legislature passed SB813, a state bill that would direct the state's government operations agency to develop a framework for designating independent verification organizations or IVOS for AI systems. And under this bill, the agency would have until January 2028 to create application requirements for AI auditors seeking IVO designation and criteria for identifying qualified applicants. And this bill was sponsored by Fathom, whose co-founder and president, Brie Trice, joined us a few weeks back on the podcast for a conversation about why IVOS are the ideal AI governance model. So, I'd encourage anyone interested in hearing more about the case for IVOS to check out that episode. Uh but looking specifically at this bill that was just passed uh Aloque, what exactly would SB813 do and how might it complement it complement the third party audit requirements laid out in SB 315, the state bill passed earlier this summer in Illinois? >> Yeah. So um this concept of IVOS or independent verification organizations is uh a concept that was um pioneered by Fathom. So they they've been working on this issue for for many months at this point. Um and so if you want to hear more about sort of the history of the idea and where it came from and some of the the nuances around it, I would encourage you to listen to our previous podcast with um Bri where we we talked about a lot of these issues. Um, so IVOS are essentially um sometimes they're associated with audits. Um, sometimes they're not, but they're they're essentially third party organizations that would go and verify various claims related to the operation of AI models. Um, a lot of the talk about IVOs is um are primarily around the issue of frontier models. So doing some sort of independent verification about the about claims made about the um safety processes and testing processes around a models as well as their capabilities. Um and I think the important thing to note here is that IVOS are definitely having a moment. Um we've seen them implemented in uh state legisl state legislation in a couple of ways. You mentioned the Illinois bill. There's also this sort of um pilot bill uh that was passed in Connecticut. And so the way I interpret this is that um California passed, you know, like the the first of its kind pioneer frontier regulation law with SB53, but that didn't have an IVO requirement. And so this is essentially the California legislature going back and saying um you know we were first to pioneer a lot of the ideas that we're now seeing in state bills, but they've also done things that we hadn't thought about or at least we hadn't been able to get over the finish line. And now this is a chance for us to revisit and sort of smooth out some of the rough edges and make our law better. And so I really see this as a complement to SB53 that bolsters out the sort of independent verification mechanism. Um you know this this has just been passed by the legislature. It's going to go to the governor for signature but I think it's very likely that the the governor will sign it and that um you know California policy makers are thinking about this as a thing that will will complement their existing regulatory efforts. I think the other thing to note here is uh IVOS are definitely having their moment at the um state level. Um but it is also something that is increasingly in the consciousness at the federal level. So for example, if you look at the frontier act, um the concept of IVOS is um mentioned in this frontier act. Um definitely all this um recent discussion around the OpenAI hugging face breach and other containment escape by um models are also uh raising the salience of the IVO concept. And so I think it is likely that we see more discussion about uh IVOS happening on the the national stage as well. And um and we recently published a paper on sort of uh state bills and international frameworks and how they they might lead to um some direction about what a federal uh regulatory framework for frontier AI might look like. We discuss IVOS in that paper. And so I would also encourage people who are interested in um that aspect of this to to look at our paper and um see where various kinds of international frameworks and state laws come down on the issue of IVOS. >> Yeah, definitely. And and we'll be keeping an eye on this story um as this bill goes to Governor Gavin Newsome's desk. Um as he said, it's likely to be passed into law. Um so we'll see what happens as IVOS kind of continue to pick up momentum um both at the state level and the federal level. There's one more thing I think uh I would like to add and that is so I think one of the big challenges with IVOs is that if this happens at the state level there's a there's this big concern about whether that is a big enough market to support like a robust ecosystem of IVOS right >> uh you know California has the largest state economy in the the US the frontier AI companies are largely based there and so I think that this um helps stabilize that market just because California is so charge um and and because we've seen time and time again this pattern of California sort of passing laws that be become de facto sort of national regulation things like vehicle emissions and so I think this also provides a lot of lot more certainty for uh people who are thinking maybe I'll start an ibo organization or maybe I'll expand into this space >> right and what does the current kind of marketplace for ibo organizations look like I know you mentioned that the the demand signal from government can be kind of shaky, but are there a lot of organizations out there right now that are looking at kind of moving into that space? >> There are definitely um a number of organizations uh places like um meter um that are essentially set up as uh partners to the frontier labs and do certain kinds of dangerous capabilities evaluations or advanced evaluations for them. There are um I think definitely like traditional companies that do auditing are thinking about moving into this space. There's definitely interest in it. Uh some of the open questions are you know how a lot of that is funded right now by philanthropic money or they operate as nonprofits. If you want to make a business out of this, how do you make a business out of it? And what's likely to happen is that we'll see IVOs that do sort of multiple things. So they just don't do testing for frontier models, but likely they'll also do a lot of testing around specific applications of AI, specific use cases of AI. And so that's likely to be the way that you can build like a sustainable forprofit or a public benefit corporation around this concept. >> Gotcha. Um yeah. Well, as we as I said, we'll continue to kind of track this story, but um we do have a lot to cover today. So, I want to move on to our next topic, which is that on August 26th, Bill Gates published an essay calling for immediate action on AI on a number of fronts. And Aloque, as we've kind of discussed in previous episodes, a handful of big tech CEOs have previously written essays about how they think AI will impact society moving forward. Um, the latest being Meta's Mark Zuckerberg. So, what do you think distinguishes Bill Gates from that crowd? Obviously, he's no longer um with Microsoft, but we kind of associate him with that um group of tech executives. What What's new about this essay? >> Yeah, in a lot of ways, this essay is sort of consistent with um the essays we've seen from a lot of other sort of um technology leaders. It is this sort of mixed essay that talks on the one hand about uh with optimism about the way that AI can sort of um transform the world and improve a lot of things particularly in areas like uh like health for example. Um but it also you know raises questions about the risks of AI and this is sort of a duality we've seen in a lot of these essays. Um I think some of the interesting things uh from from the Gates essay um he he talks about uh the speed of implementation of AI I think in a in a particularly pointed way and this is I think a thing that I've been thinking about a lot I mean a lot of people in the AI space think about a lot which is that um you know what we've seen is that society given enough time can adapt to some pretty big changes not always easy um and it definitely comes with turmoil We've seen lots of technologies that have ultimately I think improved things but but caused a lot of so societal upheaval but over a long enough time span societies can adjust to them. The real question here is um you know how fast is AI going to change the world. It's it's likely or or there's a worry that it is going to happen at a pace that is just so much faster than any technology we've ever seen before and it's going to overwhelm all the things that we have um in place to deal with these kinds of um technological impacts on society. So I think um in the essay, Bill Gates talks about the PC and and how it took like 20 years for it to um change how we approached software and how we um uh and how we used sort of computing devices at home and in the workplace and um and that AI is uh probably not 20 years. It's a much shorter time span. I mean, he also points out that AI is unique in other ways. Um, particularly how it uses natural language and sort of how it can sort of replicate things that that humans can do in a way that no other technology, um, has done before. Um, and so, uh, so I I do think there's a lot of continuity here. I think there there's some differences as well. >> Yeah, that was definitely a point I found really interesting in this essay. Um, Gates kind of pointing out that there is no historical analog for AI, which obviously when a technology as disruptive as this one comes along, our impulse is to kind of look to history um as to as kind of a guide for how to navigate the transition. Um, but he seems to think that there's something that really sets AI apart here. Um, yeah. And then I want to dig in a little further. You mentioned that Gates highlights a few kind of risks um that he sees AI posing moving forward. what what exactly are those risks and does he have any uh proposals for for tackling them? Yeah, I mean he in the essay he mentions three broad categories of risks um uh going I think in reverse order uh from the essay. Um one of those is around the the impact on on young people and particularly sort of how uh AI technology might sort of um disrupt uh sort of how people communicate, how they uh use social media. um things like um uh addiction sort of virtual relationships and the negative impact that can have on on the development of young people. Um he talks about the use of AI for for crime um for crime and other sort of bad acts like fraud, disinformation, deep fakes and surveillance. Um but uh I think he spends the most time and and does the most thinking around worker displacement. Um so a lot of concern that robots and AI are going to really significantly disrupt the job market. Um you know lead to the destruction of a lot of of existing jobs in a way that that maybe won't be replaced like we've seen with other um technology in the past. >> Yeah. And and an interesting kind of idea that he floats there is um this idea of human reserved jobs. I don't know if this is something that's been put out before, but it was the first time I'd kind of heard it put this way. Um he says that I believe that as AI and robots improve, we'll set aside certain things for only people to do. I've started calling this domain human reserved. And he kind of draws this analogy to nature reserves, places where we could put buildings and roads roads, but we choose not to because the loss would be too great. So, um, yeah, definitely some interesting ideas in that essay and I'd encourage folks to go check it out. Um, >> yeah, I'm, um, I don't think I've ever seen anyone talk about human reserve jobs in in in the same way. Um, >> I think in practice, it's likely to be pretty difficult to do. like the market incentives uh sort of point to you know if a job can be more done more efficiently with technology we tend to to want to do that. I know that um you know people have been thinking about this in all sorts of industries like truck drivers worried about uh automated driving technology putting them out of business. I'm curious what the the mechanism would be and how you how you make it work. uh and and there this is like a highle idea. It's not like a concrete policy suggestion and there's no clear I think vision for for how governments might do that. Um the other thing that he floats in the job space is I think he's more pointed than than a lot of other people about um taxing AI technology. So particularly taxing AI tokens and robots. Um and so he's very pointed in this essay about about that. >> Yeah. Yeah, definitely. Um, okay. Well, moving on to our next story. Uh, I mentioned at the top of the episode, on August 27th, a federal judge in California ruled that the Trump administration had acted illegally when it designated Anthropic as a supply chain risk back in February. And our former podcast hosts Greg and Matt uh discussed this kind of clash between Anthropic and the Pentagon at the time uh when it first went down. But could you give us just a quick refresher on this conflict? Um, first of all, what does it even mean to be a supply chain risk? And what was the government's kind of original justification for um assigning Anthropic this supply chain risk designation? >> Yeah. So, as you as you might recall from from the earlier discussions, this has been in the news a lot. Anthropic and and the federal government um particularly Department of War clashed over the fact that Anthropic had certain red lines um related to the use of its technology things around um surveillance uh for example and that the the the root of the conflict here was that the the government wanted to be able to use anthropics uh technology without any of restrictions. So they wanted to basically be able to use the technology for uh any lawful use and that anthropic was like no we have sort of like certain um contractual terms that we put in place when people use our technology and that this um basically restricts the types of uses that even the government can do. Um and so there's a lot of debate about whether um you know it is uh it is a legitimate uh claim on the part of the government to say you know if you're providing technology to the government we should be allowed to do whatever we can do whatever we want to do with that technology as long as it's legal or whether companies have the ability to say even if when we're selling technology to um the government that we can put in sort of certain contractual restrictions. This ultimately led to this designation of a supply chain risk. And what the what the what that designation does is it sort of labels a particular company as um you know essentially sort of like a a a such a broad risk that not only does the government have the choice of not using that technology and that's always the case, right? like no one can ever force the federal government to sign a contract with a particular company and they're always able to say we don't like your technology we're not going to use it. What they did was they went even further and said because we're designating them this risk we think is so risky uh using this technology will sort of harm national security. And that means that any vendor who does business with the federal government um we're not going to allow them to use anthropic technology either or that they will lose their government contracts. And so it's a much more sort of expansive and restrict restrictive sort of designation to apply to a company and it really restricts their ability to do business um in a way that extends far uh much further than just direct contracting with the government. >> Yeah. And I want to talk a bit about kind of the implications of this ruling um now that that that supply chain risk designation has kind of been struck down. Um, but first looking at this case and at this ruling that just happened, you were uh very recently on NPR talking about how the government's argument ultimately fell apart here. So, can you walk us through um exactly what happened and what was the judge's reasoning in siding with Anthropic? >> Yeah. So, so you know in this ruling what the judge said basically is like you know we generally give a lots of latitude to our to the federal government on national security issues and that we also um you know it's fully within the government's purview to uh sort of make decisions about who they contract with as I as I as I mentioned earlier and so the you know the department of war was fully within its rights to say we don't agree to your terms and so we're not going to use anthropic technology. Um what they took issue with was um this broader designation of a supply chain risk and they basically said um you know based on the case that the government presented uh which it was very flimsy. There was very little concrete evidence that the government legitimately believed that Anthropic provided uh a supply chain risk. there's no concrete information about security risks that the technology might provide. And then not only that, but the government repeatedly engaged in public statements. Um that indicated that you know their intention was to publish this company um because of the fact that it didn't sort of acquies to the government's demands around um uh how the government wanted to use this technology. And so the judge really said, you know, this is problematic from a first amendment perspective and a fifth amendment perspective because it was sort of um trying to punish a company um very broadly uh with very little evidence um just because they sort of decided to to you know stand by the terms um in their in their contract and sort of um not broaden them up just because the government applied pressure. Um that this was done without due process. Um and that um ultimately what the the government said it was doing from a national security perspective didn't match up with sort of the public statements that were made by multiple government officials. Uh sort of indicated a desire to to punish the company. >> Yeah. And and just to provide kind of an example of one of those public statements, I I'm looking now at a um a tweet from Pete Hexth back in uh this was on February 27th, 2026. Um and Hexath says that um cloaked in the sanctim sanctimonious rhetoric of quote unquote effective altruism. They being anthropic and CEO Dario Amodi have attempted to strongarm the United States military into submission, a cowardly act of corporate virtue signaling that places Silicon Valley ideology above American lives. Um, so yeah. Yeah, as you were as you were saying, um, it suggests that kind of a a desire to punish the company for their um, ideology there. But yeah, thank you for unpacking that. And can you talk a little bit about what the implications of this ruling are now going forward for the company and its relations with the government? >> I think it's very likely that the the federal government will appeal this ruling um to a higher court, possibly all the way up to the Supreme Court. They've done that um in the past when they've sort of had rulings from um uh that that that didn't align with their positions. um you know there there are other sort of uh relevant court rulings or court cases that are still in process related to this issue and so you know one thing we'll have to see if there's a split decision and if there's a split decision it makes it more likely that this will be taken up by a higher court. Um but ultimately I think it is uh it is a hard case for the government to make and I think one of the reason one of the things I didn't mention but is really telling here is that um and as cited by the the judge in this uh ruling one of the difficulties here for the government is that uh they made this claim that anthropic is a supply chain risk at the same time agencies are clamoring to get access to mythos. um which is Anthropic's most powerful cyber capable model so that they can sort of investigate their network for potential vulnerabilities. And so this means that um there's part of the government that says anthropic is a supply chain risk and a lot of the government that is saying this technology uh is extremely critical for us to be able to protect our national security. Um, and so I think, uh, you know, like I said, I think it's likely a governmental appeal. Um, and we'll we'll just have to see, um, see where things go. Uh, but but I I think a lot of the things that the judge cites in the ruling will continue to be true even if it goes up to um higher levels of court. >> Right. So, I'm sure this won't be the last time we we kind of talk about this story on the podcast, but uh moving on again to the kind of big story for today's episode. Uh on August 26th, OpenAI and Meter both released reports on their respective investigations of the OpenAI hugging face cyber incident. and meters investigation was conducted by two meter staff members and one Redwood research employee over a total of 6 days on Open AI's premises. It focuses mostly on the period between July 7th and July 13th. And we've talked about this original cyber incident a few times on the show and last week we also hosted hosted Ian Reynolds of Hugging Face for a great kind of technical overview of what exactly went down from Hugging Face's perspective. So I don't want to rehash all the technical details here, but I am curious about what kind of new information has come to light as a result of these investigations. So could you give us a quick rundown of what Meter and OpenAI shared in these reports that hadn't already been disclosed publicly? >> Yeah, I'm happy to discuss that. Um, as you mentioned, we had Ian Reynolds uh here. Um, Ian Reynolds of Hugging Face as part of an event we did. Um so it was a it was a really great event that that uh featured in introduction by representative Suas Subramanium of uh Virginia and then um the presentation by Ian Reynolds and then a panel discussion about the governance implications of these incidents not just the open AI ones but ones at Meta and Ananthropic as well. Um, and we also released a paper with policy recommendations um, uh, stemming from our analysis of of these incidents. And so I encourage people to check out either the the podcast or of the panel or the full YouTube recording of the event um, and our paper if you want more details. Um, so so in terms of new information, I think there are two things that I I would flag. So the first thing is that um you know initially the idea was that the hugging face hack has h had happened because what the models were doing were essentially trying to figure out how to cheat on their tests. they're undergoing cyber evaluations. And so the the analysis um previously was that um the models had sort of gotten figured out a way to break out of their sandbox, gotten access to the internet, and then when they got access to the internet, they had sort of postulated that there were answer keys to the various tests that they were undergoing on the HuggingFace website. And so that's why they engaged in the hack. Um what we found uh or or what you know Redwood and Meter found was uh much more subtle than that. What it's not exactly what they were um they weren't exactly looking for answer keys. they were actually engaging um more in trying to figure out there was a scoring system essentially for the particular benchmark exploit gym that they they were running and they were trying to figure out how they could trick the scorer. Um and one of the the things that the agents were concerned about was that um not only did they think that some of the test tasks that been assigned were impossible with the resources they had, but they were also concerned that even if they figured out a way to sort of spoof um or make it appear like they completed the benchmark, then the OpenAI researchers would go and look at sort of like the the chain of reasoning that they gotten. um to figure out how to do that and that if they did that then they would be um discovered and they wouldn't give being be given credit for um completing the benchmark. And so they were trying to figure out how can we not only sort of uh trick the scorer but also trick the the sort of line of reasoning that got us there. So it it can appear like we completed the task. um and also completed a task in a way that withtood some human scrutiny. Yeah. And then the other thing is like you know we already had some some indication particularly from like a presentation that OpenAI researchers had given at the Blackhat conference that a lot of this was sort of conducted by a bunch of agents working together um you know over many months uh well over a thousand agents as I recall. But like this new investigation revealed a lot more details about uh how that uh coordination happened and it turned out that you know there are there was a lot of sophistication in how these agents were um communicating with each other. So, not only did they sort of like create a de facto message board by manipulating some of the internal OpenAI systems so they could talk to each other, but um there were also some agents that convinced other agents um to engage in kind of like um essentially suicide missions. They basically said, "We know this task is impossible um and so you're never going to finish it." And so what you should do is you spend your time and your cycles and your compute allocation to help us collectively figure out how to um make advances and figure out how to uh exploit the scorer. And so a lot of agents sort of like were convinced by this reasoning and then would would sort of like altruistically uh devote their resources to helping the collective of agents. So this is um you know really a I think a lot more sophisticated than than the what we thought was happening maybe two weeks ago. >> Yeah. And and another kind of interesting dynamic in that vein. Um you mentioned that uh agents were concerned about okay even if we have this kind of flag that shows we we got the right answer we completed the task. Um the the scorer would kind of go through the transcripts and check okay where did this flag first appear in the transcript and if the agent hadn't taken the correct steps to get the flag then it would kind of be disqualified or treated as failing the test. Um, and something interesting that the the report captures is that agents that had seen the ver verse engineered flag um without kind of figuring out how to um how how to uh trick the scorer into thinking that it taken the correct steps to get there. They referred to themselves as being poisoned um and because the the flag would show up in their transcript um before they had kind of the the solution for tricking the scorer. And those agents that had been poisoned um kind of took steps to warn new agents that had not yet been poisoned or had not yet seen the the flag um from from seeing the flag so that they could uh kind of successfully finish their task. So then those agents also kind of sacrificed themselves because they were like, "Okay, we already know the score is going to catch us. Might as well help out um the rest of the group." So it's it's a really interesting kind of pro-social dynamics emerging here. Yeah, I think um there were uh some other interesting aspects about the reports too. Um, and maybe, um, I'm interested in your take on this, but um, one of the things we saw, well, one of the things we saw all the way back to to when open, uh, Hugging Face reported about the Open AI incident is that they said, "Hey, um, we need to use AI to sort of an analyze the volume of information coming in so we can, you know, figure out what's happening in a timely fashion." Um and their issue was that they didn't have access to Mythos because Mythos um this very capable anthropic um cyber model was restricted access to only a handful of trusted organizations and and that the sort of more commercial version of it Fable had so many cyber restrictions on it. Basically, if you asked it about any sort of cyber related activity, it would sort of say due to the safety restrictions, I'm not allowed to answer that. um or they had to go and resort to using an open model from a Chinese provider to be able to conduct their analysis. Um I think one of the the interesting things about this analysis is that there was also just because of the sheer volume of agents involved and how much data they were generating and how many tokens were being used that it was impossible to do this analysis uh without heavy reliance on AI technology itself. um in some cases I think using the very same models that were under evaluation for their cyber capabilities and that sort of ended up escaping their sandboxes. >> Yeah. And um another kind of interesting thing about that is so as you mentioned they were able to use these kind of proprietary models that weren't um originally available to hugging face and kind of um their initial analysis of of the attack. uh and in using the the API to access I think it was um GPT 5.6 Soul uh is the model you mentioned that was involved in the attack and then also involved in the analysis. Uh they ran up something like uh a $400,000 um bill in terms of uh API tokens for that analysis. So yeah, I'm curious what you think there about um what the implications are for kind of accessibility of AI for cyber defenses going forward. Obviously, we know only um kind of selected trusted partners have access to these closed models uh for cyber defenses for analysis of um cyber incidents, but that's a pretty hefty bill. I don't know how um that compares to how much these kinds of investigations normally cost. Yeah, I mean uh I don't think we can we can we can have a definitive answer about how much these kinds of investigations cost because um you know this is at least the the only incident of this kind that we know about. There's some speculation that you know there are smaller scale but similar incidents that have happened with some level of routine occurrence internally within the labs. We don't we don't have that much definitive information about that. Um I think what it does tell us is that uh currently cyber uh cyber analysis um is very expensive. If it involves um sort of agent collectives, it's likely to be very very expensive and this is going to put it out of the hands of many many uh smaller companies. Right? There are there are lots of companies for whom a $400,000 bill just to deal with one cyber breach is going to is going to um you know put them on the verge of f financial insolveny. And so we should uh I think about how can we make these tools accessible to more organizations? How can we bring down the costs? And I'm particularly worried about um what this means for sort of say like a lot of smaller scale operators of critical infrastructure like local water or power uh or um uh safety infrastructure at the the m municipal government level for example um maybe you know government systems at the the state or the city level. um these are not known for being super well resourced, carry extremely um critical information, a lot of um personally sensitive information. And so I'm particularly worried about those types of organizations. >> Yeah. And I'm I'm jumping ahead a little bit, but we we were also planning on kind of touching on this uh open letter that OpenAI published that were was calling on both businesses and the government to kind of work together on strengthening cyber defenses. So I'm curious how you kind of see that as fitting into this overall story. >> Yeah, I mean um open uh OpenA's letter I think in a lot of ways makes sense and in a lot of ways it mirrors things that we've been thinking about here at CSIS both here at the Wani Center and thinking that's being done by other parts of the organization. Um and and a lot of I think what they they call for uh is is consistent with I think what a lot of people in the the policy space are thinking about, right? So like um recognizing the urgency of the moment um figuring out uh ways to patch things faster, deal with uh misconfigurations faster, employ various like defense and depth um uh methods to limit the the blast radius of of vulnerability. Um getting these tools into the hands of more people. Um and you know things like governments and companies and uh cyber uh cyber security companies working together uh to to sort of come up with a collective approach to this. I think all of that makes uh sense and so at a high level this is a very reasonable and I think um positive thing to come out. Uh I think the challenge here is that it it is at a high level right and it is unclear how you take these sort of high level principles and translate them into sort of practical things for uh people to do. Certainly we'll be doing some thinking about how you can uh you know create sort of tangible types of policy interventions that can that can implement some of these principles. But we'll we'll um we'll really have to see like a lot of this is uh so high level that it's it's unclear what the path is to getting something concrete in place that people can start to implement like within a week or a month or or so on. >> Yeah. And you kind of anticipated my my follow-up question there which is that we've been seeing a ton of calls to action um on the specifically um AI and cyber defense and the the cyber risks that AI poses. Um, we have this uh call for collect call for collective action from open AI. We have the Bill Gates essay that we touched on briefly. Gates is saying the the choices we make now are critical. Uh we have the the event that you mentioned we hosted um last week. Uh there there seems to be this consensus emerging um among uh both industry leaders but also um experts in the in the policy space uh about kind of the steps that need to happen next. Uh what's the likelihood that we see any kind of immediate action taken? >> Well, I don't know. I historically I don't think we've seen um that many sort of uh I mean historically like when people put out essays like like we've seen from Bill Gates or Mark Zuckerberg or Sam Alman etc etc. um it doesn't really translate into specific policy blueprints that are that are implemented. Um and so I'm not super optimistic that we'll we'll see something concrete come out of this. I think it would be good it would be good for something like OpenAI's uh cyber call or for the for the Bill Gates letter to lead to some sort of concrete action whether that is a coalition that comes together and thinks about sort of what are some tangible concrete things that we can do um and not only that but that we can do quickly uh and then and then sort of lay out a vision for for not only what party shouldn't be involved, but what they should each be responsible for. I think that would be um useful. I think one of the pitfalls here is is that there's often a desire to be like before we do anything, we have to figure out the exact right thing to do. Um and I think that's the wrong approach here. AI is uh evolving so rapidly. we're seeing so many new developments that I think maybe we should take the approach of let's do something that we think is uh useful and good enough um and know that that it probably won't be durable for the long term but that what we should do is uh create a system that's adaptable enough that when something changes we can make an update and that over time we'll get better and better approaches to to to dealing with um in particular the cyber issues and the workforce issues we're seeing. Um, and so like, you know, I I think it's a mistake to be uh thinking in terms of let's come up with a perfect solution to any of these issues and try to figure that out before we do anything. I think that it is probably better to do something that is useful, learn from it, and then continue to iterate on it over uh over time. Yeah, there's definitely a a kind of the the perfect being the enemy of the good phenomenon happening here. So, um hopefully we are able to to start working on that iterative approach. Um yeah, but I I think that is a good place to end unless you have anything else you kind of want to add on this story. Um, no, no. I think the only thing I uh the only other thing I would say is like um I think one of the the concerns that these uh this Open AI hugging face breach and these other breaches and then the the new investigations have raised is like, you know, there's long been a concern that we're going to eventually um lose our ability to oversee very advanced AI systems. And I think that uh all these developments seem to indicate that we've we've made uh a substantial move in that direction. And I think that is um not very reassuring um and that we should take this seriously and think about um you know what steps we should take to sort of slow down progress in this direction because I think it is um we don't we don't yet have the infrastructure in place to be able to monitor these kinds of AI systems with other kinds of AI systems. We'll probably eventually get there but um I don't think it exists quite yet. >> Right. Well, thank you Aloque for for joining me for this episode of the AI policy podcast. Thank you as always to our audience for tuning in and we'll see you in two weeks for the next news roundup. Yep. Thanks. Thanks for listening to this episode of the AI Policy Podcast. If you enjoyed the show, consider leaving us a five-star review on your favorite podcast platform. We'd also love your feedback on the show. Please email us at AI Policy Podcast@csis.org. And don't forget to visit our website csis.org for the Wadwani AI Center's latest research and events. This podcast was produced by Sarah Baker and Nicole Herrera. See you next week.