Anthropic adviser says it's "not hypothetical" that AI could abet biological weapons risk

[0:00] should the government regulate ai and if so how we're joined by cbs news cyber security contributor [0:05] chris krabs who ran cissa in the first trump administration and ben buchanan who advised [0:11] president biden and is now a professor at johns hopkins and an advisor to anthropic good to have [0:16] you both here good to be here thanks for having us so this was a big change because uh president [0:22] trump and his advisors clearly saw something that led to a shift in their policy they were very light [0:27] touch literally don't do anything and now there was this executive order just a few days ago [0:34] to have some step towards i don't know if we can call regulating it because it's voluntary but that [0:41] the companies have to voluntarily provide their very sensitive technology 30 days before releasing [0:48] it to let the government have a first look what does this accomplish chris well i think the biggest [0:54] concern was while there had been a sense that generative ai and some of the more advanced tools [1:00] were giving cyber attackers and defenders better capabilities it wasn't until about two months ago [1:07] where anthropic released their mythos or at least announced their mythos preview that really brought [1:12] it i think to the forefront with cabinet secretaries like scott bessen and treasury and some other folks [1:19] so they they became more concerned they go to the white house say hey we need to take a bit more of [1:24] a proactive approach and so they got concerned because business entities yeah i mean in part [1:29] the trade the finance sector banks jamie diamond and jp morgan and a number of others were like [1:34] whoa whoa we need to pause we need to think about what we're doing here and we need to take a bit more [1:39] robust security focused approach on how we uh as this avalanche of vulnerabilities that will be discovered [1:48] by ai how we prepare uh in and ensure that we're not going to be overwhelmed ben 30 days for the [1:57] government to conduct a full security vetting is that possible or is this just sort of window dressing [2:04] to say the government's doing something in the viden administration we built a capability at the ai [2:09] safety institute to do quick uh voluntary testing of ai systems including for cyber risks and for biori [2:15] so i think the tests themselves can go quickly um but that doesn't mean the government is going to [2:20] do that in a fulsome way so the the jury i think is still out on how the president chooses to implement [2:25] this right but the the timing here i think is is not one of the pieces that worries me it's more [2:29] getting the details of what does this testing look for right and what's the consequence exactly so [2:34] like can the government stop it the whole risk and worry had been this would get in the way of america's [2:41] arms race with china does this get in the way of america's arms race with china not as it's currently [2:46] written no i don't think so and i don't even think even more stringent versions would i do think what's [2:50] happening here is you see the government wrestling with something that we wrestled with in the biden [2:53] administration this technology is coming from the private sector and in many respects it's the first [2:58] revolutionary technology that is primarily in its current form from the private sector you think about [3:03] nukes or space or so much the government has and has had a role in all those technologies [3:08] inventing them shaping them it doesn't really have that in ai so the trump administration is now [3:12] trying to figure out how does it want to assert itself in this new domain that actually is quite [3:16] different conceptually than many of the ones that came before it and in congress they haven't [3:22] regulated really not yet in fact the executive order is very clear when it says this is not a [3:27] licensing regime there is not a pre-clearance mechanism but it does set up a structure where [3:33] the nsa the national security agency can define what a covered model is and then brings those [3:40] models into a the 30-day pre-review and then you can look at maybe what some of the vulnerabilities [3:46] that could be discovered and then you can share those out with defenders with industry so that they [3:51] are prepared but separately this week you had congressman obernolte introduced draft legislation that may [3:59] give a requirement for submitting some of these subsequently covered models for pre-review and [4:08] auditing by independent bodies on a twice a year basis unclear if that's got a future though but it's [4:16] starting the conversation which i think is critical so the president did say to reporters that all the [4:22] big ai leaders are coming to the white house quite soon he has said he's looking into the u.s government [4:28] taking an ownership stake in some ai companies that mean us u.s taxpayers would own stakes and [4:34] it's been reported that open ai sam altman also pitched an idea to senator bernie sanders [4:40] about transferring 50 of the equity of the top ai companies to a public fund to both of you would [4:48] that work what does that do is that just retrofitting what you said which is the government needs to have [4:52] some kind of hand in here it's certainly conceptually trying to adjust address that question of what's [4:57] relation between the public and private sector i think with something that is that drastic the devil [5:02] is really in the details and we don't have any details of this meeting or even of senator [5:06] sanders proposal so i think we've got a long way to go on making that real and it feels incongruous [5:11] with me of uh what does it mean to let the private sector run here so i think there's a lot of [5:16] complexity that's yet to be worked out in that proposal and on that note it also creates a significant [5:21] governance challenge you know when you control when you own part of uh whatever the tool is it creates [5:27] a uh a conflict of interest and some self-dealing perhaps if you're also trying to regulate it so [5:33] again the details here are what's going to matter um it is interesting that it's almost a horseshoe effect [5:38] where you have both republicans and democrats that are talking about you know perhaps we need to take a [5:43] take an ownership stake in these models they seem scared i mean that seems to be the takeaway that's [5:49] forcing a change in position and i mean anthropic the company you advise called for an international [5:56] pause in ai development because of how fast the systems are improving as i understand it the [6:01] concerns the models are becoming so advanced they can fix themselves without humans being involved [6:07] well the dynamic here that's very interesting is what we call recursive self-improvement so the ai [6:11] systems help the ai companies generate the next version of the system and that can show up in a number of [6:17] ways one of the ways it can show up is the ai systems write the computer code for the next version [6:22] another one is the ai systems do the research and some of the math and underlying conceptual work that [6:28] leads to efficiencies in the next system and now we increasingly see ai systems helping design the chips [6:32] that are the lifeblood of this so i think it is the case for a variety of reasons recursive self-improvement [6:37] massive infrastructure investments we should expect these systems to get continue to get very good [6:42] very quickly and that does raise real national national security questions no matter who's [6:46] president do you expect that most companies are going to comply with this voluntary 30-day review [6:53] chris well i think there's a separate mechanism that probably has more teeth and that's what was [6:59] issued on friday a national security presidential memorandum on the national security enterprise [7:04] in the use of ai which allows the department of defense to use the power of the purse the [7:08] procurement mechanisms to dictate what the policies are and who needs to submit to reviews and who [7:15] will you know follow certain presidential or rather department level direction so i think the executive [7:20] order is an interesting framing for a government-wide approach but the real movement i think the real [7:27] teeth is what's happening at the department of defense uh in dictating what the policies are for how [7:34] vendors can uh dictate to the government on how models can be used in the biden administration in [7:39] 2023 we brought the companies to the white house to meet with the president and they all made voluntary [7:44] commitments which they then followed through on to do independent red team testing for risks like cyber [7:48] for bio to publish their testing results to the public and they follow through on that so i do think [7:53] there's a history here of companies working with government but i suspect both sides are going to have to [7:58] do a lot more as this technology gets better i recall during the biden administration there was this [8:04] agreement between the us and china not to let ai touch nukes that's right as i understand it there's [8:10] discussion about the risk towards biological weapons a open ai anthropic and google's ai initiative signed [8:17] a letter saying that there should be a law to keep ai out of bio biological weapons chris spoke very well [8:24] and you spoke very well about the the cyber dimensions of the current set of models i would look at these [8:28] current models not as cyber models but as generally capable models and the bio what does that mean sorry they can [8:34] do uh expert tasks in a wide variety of areas not just in cyber also in bio and the like this was [8:40] something that we were very alert to in the biden administration there's a long section on bio in [8:44] the president's executive order in 2023 there's upside you know ai can do a lot for medical discovery [8:49] but there also is legitimate concern that ai is outperforming phd level virologists on virology [8:55] questions and that could abet a bioweapons risk that is not hypothetical okay well now you're scaring me [9:01] i mean do they need to be bilateral agreements between countries or this is because we have the [9:08] edge at this moment america needs to set the rules the world for everyone i mean essentially they did [9:12] for nuclear weapons right we had we had them first but it was out of fear look ai emerging technologies [9:21] quantum is going to present i think the same problem but we're now entering this governance space [9:28] where technology development is moving faster than democratic oversight and control can allow or [9:35] provide and i think to your earlier question about they're scared there's fear it's it's they're [9:39] overwhelmed they don't really understand the technology it's very complex uh understanding how [9:47] to meaningfully intervene that's consistent with the american regulatory tradition of its light touch free [9:53] markets capital markets uh we're seeing a bit of i think a divide here on how to interact but again [9:59] you are seeing some interesting parallels in direction of travel that's consistent uh between the [10:05] administration the congress and democrats and republicans uh that that understand something needs [10:10] to be done we're just not quite sure what the mechanisms are just yet yeah and the technology is [10:16] moving really quickly while that's being debated a gentleman thank you for bottom lining this for us we'll be right [10:22] back in a moment