00:00Let's bring in Ali Mellon for more, principal analyst at Forrester and also author of Code War,
00:05How Nations Hack, Spy and Shape the Digital Battlefield. This is so interesting because
00:13Anthropic has basically become, Ali, the case study in the first instance of how the different
00:21arms of government have oversight of the most powerful frontier models. Would you just reflect
00:27quickly on the news that Anthropic is allowed to proceed with re-enabling Fable access to entities
00:35around the world? Absolutely. Well, first off, thank you so much for having me. This is a very
00:40big win for Anthropic, of course, and also for the U.S. government because the reality is we need the
00:45space and time to innovate, especially on something this important in these frontier models. And
00:52hopefully this is a good way for the U.S. moving forward to understand exactly what needs to happen
00:57to make sure that we are protecting national security while also enabling innovation. Because
01:01of course there are two sides to national security. There's the element of protecting us against what
01:06could attack us, but there's also the element of making sure that we have the defenses to do so,
01:10which requires using these models. I think it's really hard to understand some of this.
01:17The jailbreak concern is that actually there are guardrails in place to ensure that a malicious or
01:23bad actor doesn't use the model for malicious purposes, right? But banning outright doesn't
01:35kind of solve the problem totally. You're an expert in this domain. Just trying, in layman's terms,
01:40explain that concept. Unfortunately, the reality with models like this is very similar to what we see
01:47with software. There's always going to be bugs and vulnerabilities that we just don't know about
01:51yet in software because we can't test every single variable in every single way. The same is true for
01:57some of these models. There are different ways of asking questions to these models any way that you
02:02can possibly think of that are going to change the output. And so while Anthropic and other model
02:08providers have put safeguards in place to try to limit that and limit the risk as much as possible,
02:12we have to accept a more dynamic reality where there are going to be prompts that we don't expect
02:18that are going to give access that we also don't expect. Now, Anthropic is doing a lot to limit the
02:24potential damage of this, including making sure that they're monitoring the prompts that are coming
02:28in, that they're putting very strict guardrails so that even if there is some access that's gained,
02:34it's very limited compared to what could be possible and compared to what could be dangerous.
02:39But it takes a balance of controls. It's not a black and white issue where we can just outright
02:45block access and have no problems. We, to a certain extent, are going to continue to experience
02:50issues like this. We just need to handle them in a much better way than outright blocking access
02:54and causing challenges for organizations. Industry wants certainty. In your analysis of the government's
03:03behavior and action so far, how do you assess the level of certainty that we will or won't get on
03:11policy? Unfortunately, this was not the best instance of establishing certainty and trust between
03:18the U.S. government and a lot of these model providers and organizations. One of the things that we've seen
03:23is that organizations have started to try to diversify the models that they're using to prevent an issue
03:29like this from affecting their organizations again. But I will say, especially Anthropic has put out a lot of
03:35great information around this, about how they're working very closely to give pre-release government
03:40access to these models, to do information sharing on when jailbreaks happen and the severity of them,
03:46to provide dedicated resources to work with the U.S. government. And these types of playbooks are
03:51what's going to help make sure that there is certainty in the future of potentially less limited access or
03:57certainly for a shorter period of time than we saw in this instance. There are going to be bumps in
04:02the
04:02road, as there is always with innovation. But it seems like moving forward, they've at least
04:07established some common ground to move forward in such a way that organizations should face less
04:13of a hurdle in the future.
Comments