The Anthropic Fable 5 Shutdown: The AI Risk No Business Planned For

Speaker 1 00:02

This is CyberSound, your simplified and fundamentals-focused source for all things cybersecurity.

Jason Pufahl 00:11

Welcome to CyberSound. I’m your host, Jason Pufahl, joined today by Dylan Marquis.
Hey, Dylan.

Dylan Marquis 00:18

Hi, Jason.

Jason Pufahl 00:19

So Dylan, he’s been on a handful of times, but for people’s reminder, he’s the lead of our pen testing practice. And honestly, I think anytime I think AI expertise, I think, well, we should have Dylan on. So that’s a little bit what we’re doing today.

So I appreciate you covering out a few minutes. Absolutely. And actually, we’re recording this a little bit, I think a little bit more urgently feeling than sometimes because I wanted to chat about the Anthropic, I don’t know, mishap isn’t the right word, right?

But Fable 5 was released, give or take a week and a half ago. Essentially, overnight, it was globally pulled. And there’s kind of a lot of commentary around, well, why did it get pulled? And what’s the impact on business and things like that? So I was hoping today to spend a little time on discussing the release of it, then the subsequent takedown. But then I think segue a little bit into the business impacts of AI, right?

Because we always talk about the benefits, but some of the potential business impacts and really, how do companies potentially protect themselves against that, mitigate that risk? So I think a two-part, a little bit of a two-part discussion. So I know you spent a bunch of time on looking at Fable 5, and I think you understand a bit more about why it was yanked down. So maybe you can spend a couple of minutes on that.

Dylan Marquis 01:53

Yeah. So, I mean, granted, a lot of this is a little nebulous and unknown as to the actual reason and drivers that the government had for applying what is essentially like a novel usage of export control authorities against like a hosted Frontier model. So it’s kind of new territory. There’s definitely, I’m sure, going to be some legal back and forth, at least a little bit, and probably some feeling this out from kind of both angles, from the government and from the Frontier model providers. But essentially what occurred, there was some preamble to this where at least one security researcher was able to jailbreak the model. So jailbreaking is getting an AI model to do something that’s like to bypass restrictions in it.

So creating malware or giving directions on creating explosives, something that would be illicit. And that did occur. Obviously, Anthropic contests that surrounding what occurred with the security researchers said it wasn’t really novel or anything too profound, let’s say. But then kind of alternatively, what Anthropic published about this actual ban of Mythos and Fable was that they seemed to find another jailbreak themselves that they were concerned about. Again, Anthropic contested, but apparently there’s some ambiguity surrounding what actually occurred. Anthropic feels that there are kind of minor vulnerabilities. The government feels that maybe it’s something more profound. And then additionally, there’s been some talk about the NSA utilizing Mythos, and it was apparently very effective in finding flaws in its software and system. So it’s a little bit unclear as to kind of why this occurred, but what did end up actually kind of being implemented was export restrictions. So laws surrounding the distribution of Mythos and Fable, which are essentially a hosted model. So access to the hosted platforms was controlled and cannot be distributed to foreign nationals or to other foreign countries. Anthropic didn’t quite know how, can’t really differentiate who’s who. So it got banned for everyone. And this actually even extends to their own employees that are foreign nationals. So it’s pretty interesting ground.

It’s definitely something new. So I think we’re going to have to stay posted as kind of what the future of this looks like or what the next steps look like. But certainly some interesting discussions.

Jason Pufahl 04:36

One of the things that I feel, in my opinion, it should have made the headlines a little bit was, this was an export control ban, and it essentially then it applied to everybody. So it wasn’t like a half a dozen companies that have a huge spend with Anthropic were allowed to keep Fable 5 in place. It was globally pulled. And it definitely raises the question with kind of the rapid development of AI. And in a lot of cases, the rapid adoption of some of these tools, there’s always this race to have the next great agentic capability or the next great sort of LLMM quality. When something’s pulled like this, any development you’ve done, all of a sudden you can’t access those features.

So there’s a real productivity and potential business interruption. And I think as much as there’s a story around pulling it, I think there’s a little bit of a cautionary tale around hitching your wagon to some of these models and being all in, because there is some risk still, right? It’s developing so quickly.

There’s risk that features are deprecated, that features are just changed enough that it actually impacts workflow. And I think people have to give a little thought to how they implement and use these tools.

Dylan Marquis 05:55

Yeah, absolutely. I mean, the thing is that we’re kind of reading about Anthropic because of the first through the door here with a model that has these implications. Now, researchers have stated that the capabilities are not completely novel, that this isn’t something that’s totally out of the blue.

But it does feel, I mean, having used Fable and kind of the, let’s say the anecdotal kind of descriptions from others, I mean, it feels more capable, definitely. So this is going to keep happening as more and more capable models get released. There’s going to have to be a back and forth between Frontier providers and the government. And until things get solidified as to what the future, what this will look like in a mature model, we’re going to keep getting models pulled. And certainly there’s going to have to be plans to fall back, kind of concerns from consumers and kind of taking precautions to ensure that they’re covered in the event that one of their models gets pulled like this. There’s definitely some options.

I mean, things like model routers, selecting multiple providers, ensuring that your processes don’t rely singularly on one model, and probably using models that have been vetted and have been around for a while, certainly for more consistent product delivery would be like a good first step. But certainly some of this infrastructure and some of these technologies are still in development. This is kind of the new Frontier, and I think we’re all kind of struggling to keep up.

Jason Pufahl 07:31

Well, actually, I’m going to ask you to define a couple of things quickly. So Frontier model and model router. So you sort of said both of those things, and I suspect there’s some folks who aren’t familiar with either of those terms. So Frontier model, I think, seems straightforward enough, but if you can spend a minute on both of them.

Dylan Marquis 07:49

Sure. So Frontier model is a model that’s provided by a major model provider, Anthropic OpenAI. There’s some other smaller ones that also provide open source models like Mistral. So these are the hosted models that are, you know, very large, very capable, and services are provided through a company rather than an open source model like LLaMA or DeepSeek, something like that. In terms of model router, certainly this is new tech. I really haven’t even played with them myself, but I was at least familiar with the term.

So it’s a way of providing access to multiple models for a single workflow. So Anthropic isn’t responding, or Anthropic is having an outage, let’s move over to OpenAI, or let’s change out the model that’s being used. Things like that is kind of what I understand the context is for it.

Jason Pufahl 08:49

So, I mean, will that reduce the risk? If you’re heavily reliant on the capabilities of one model, and that’s unavailable, and it changes, you still may have the same sort of, potentially, right, application development, output challenges, or response issues, things like that.

Dylan Marquis 09:11

Yeah. Unfortunately, you’re kind of stuck. Certainly, I think with Fable, there was a lot of talk about how capable of a coder it was. If you’re now suddenly relying on that, and Fable gets pulled like it did. Yeah. I mean, right now, there really isn’t any competing model with Fable as of this second. So, yeah, there’s not really a good backup solution there.

Jason Pufahl 09:36

Do you feel you get much more stability out of an open source model? If people rely on those?

Dylan Marquis 09:44

You have more control. Are they as capable as Frontier models? I haven’t seen one yet that is. I know that there’s a few that are kind of burgeoning right now. I haven’t personally played with them. But, you know, to be determined, but I’d say as it stands now, the Frontier models, especially, they’re cornering capabilities like Fable. They’re the only ones that are providing that level of capability.

Jason Pufahl 10:09

Yeah. And I think, you know, I’m sort of focused on this a little bit in partly because, you know, a lot of companies that we work with, I think, are struggling to figure out how do they utilize, how do they take advantage of AI? And how do they take advantage of it centrally?

Because I think we definitely see, you know, companies where every employee has adopted their own favorite AI platform. So you get that AI sprawl and they’re using it for the more sort of common use cases, you know, helping to draft documents, doing some research, things like that. But I think it’s when you actually integrate it into your business processes more centrally that you have to really think about, you know, what model are you using, what platform you’re using, and then, you know, frankly, right, how current the model is that you’re using when you’re when you’re building and developing around this.

Dylan Marquis 10:59

Yeah. I mean, not to say that using it in an ad hoc, you know, for day-to-day tasks isn’t disruptive when a model goes down. You know, getting used to the intricacies of the way that certain providers kind of handle things, all the interface, it’s not trivial. But yeah, I would say that actually basing infrastructure and, you know, application capabilities on AI, it’s much more impactful when you lose, you know, when a Frontier model either has a disruption like what happened with Fable or, you know, probably more commonly just interruptions. These things are pretty sought after. There’s a lot of companies using them.

They’re growing very rapidly and obviously mistakes are bound to happen, you know, and outages occur. There certainly are different options. Some are hosting frontier models. Amazon Bedrock is an example where you kind of, they’re getting predefined and deployed infrastructure. Again, I haven’t personally played around with Bedrock, but it seems like a more stable option for kind of basing, building your own infrastructure against a Frontier models like that.

Jason Pufahl 12:08

Yeah. The, you know, when you look at the, if you use Claude or Anthropic, right, and you’re kind of in their interface, I think it’s interesting that they still have that, you know, sort of bolded disclaimer there that, you know, Fable 5 is currently unavailable because it’s been, I think, you know, at this time, I think it’s been about 10 days since they pulled it. It was literally out for, I mean, 24 hours at most. And I don’t even know, frankly, that most people, you know, relied on it heavily or used it heavily and yet, right, and yet they still are basically saying like, hey, you know, it’s unavailable. So it’s an interesting, kind of an interesting choice of to keep it so prominent.

Dylan Marquis 12:51

I think there’s probably a few prongs to that. One is likely the political friction that’s occurred between the U.S. government and Anthropic. I think that they’re highlighting that intentionally. Maybe not, but it certainly feels like it from an outsider’s perspective. I mean, the other thing too is it’s all good hype for them, right? No publicity is bad publicity. It’s always good. I mean, this is a big boon for them, even though I’m sure that internally they have a lot of, it hasn’t been desirable. I think at least from an advert, you know, advertising their brand standpoint, it is. So it’s, I think a little bit of it is, is just sort of propagating the hype or keeping the hype. Yeah.

Jason Pufahl 13:32

Some marketing around it.

Dylan Marquis 13:34

Yeah. Yeah.

Jason Pufahl 13:34

Any press is good press.

Dylan Marquis 13:36

Yeah. They’d be unwise not to capitalize on it. I don’t blame them.

Jason Pufahl 13:40

Yeah. Yeah. And it’s funny because, you know, while I didn’t use it, I look at it each time, I’m like, all right, I hope it’s available soon.

I want to take advantage of that.

Dylan Marquis 13:48

Same. I’m waiting for that to go white so that way I can click on it and.

Jason Pufahl 13:52

Yeah, absolutely. Give it a test.

So the, you know, partly it is interesting that it got pulled the way it did. I think our, kind of my takeaway or our company’s takeaway a little bit, you know, remains consistent with the old, you know, let your, your features bake in a little bit before you really start taking advantage of them, right. To sort of work the kinks out, work the bugs out. I think the same philosophy applies with regards to AI, which is super useful tool. Definitely want to be using it, but maybe not the most current model all the time. Right.

Dylan Marquis 14:34

Yeah. Yeah. We’re all, I mean, the, the, the Frontier models and really, I mean, I should say we’re all kind of on the bleeding edge of this. It really is moving very, very quickly. And, and again, like on unknown consequences do occur regularly. So in this kind of environment, so it’s, it’s something to kind of plan for. And I think, yeah, picking models that are a little bit older and more vetted and, and, and well-known or is a good stance at this point in time.

Jason Pufahl 15:02

Yep. So well, hey, I mean, I appreciate you jumping on for a few minutes. I know, I know I pushed to do this quicker because I, because at some point that, that, that bolded Fable 5 is unavailable is going to go away. Right. So it won’t be as relevant. But I think it does highlight the risk that’s out there. And, you know, this one in particular, you know, Anthropic certainly pushing their capabilities. And I think it’s the, really the, you know, the first big instance of, you know, sort of the government stepping in and saying, this is, we really got, we’ve really got to pay attention to this, you know, this particular release. So.

Dylan Marquis 15:36

Yep. Absolutely.

Jason Pufahl 15:38

Cool. Well, thanks. And, and, you know, as always we’ve, we’ve done, I feel like you’ve done a lot of AI content in the last, in the last basket podcast, but if anybody has questions or wants to discuss, you know, generally how they’re using AI from a security controls standpoint, or, you know, just thoughts about different models or different options that are out there, we’re happy to chat and give you some guidance and, and go from there. So as always hope you got some value out of this short episode and Dylan, thanks for joining.

Dylan Marquis 16:05

Absolutely.

Speaker 1 16:06

We’d love to hear your feedback. Feel free to get in touch at Vancord on LinkedIn, and remember, stay vigilant, stay resilient. This has been CyberSound.