A "Summer Fling": How Anthropic's Best Model Got Pulled
This bonus episode covers the abrupt shutdown of Claude Fable 5 (alongside Mythos 5), Anthropic's most powerful model, which lasted only three days before the U.S. government forced it offline. Host Andrew is joined by Thomas Moen from Norway, and the two unpack what happened, why it matters, and what D2C operators and agencies should do about it. The framing is blunt: if you wired anything into Fable 5 during its brief window, this episode is your fire drill.
The hosts walk through the timeline. On June 12 at 5:21 PM Eastern, Anthropic received an export control directive from the U.S. government ordering it to suspend access to Fable 5 and Mythos 5 for any foreign national inside or outside the country, including Anthropic's own foreign-national employees. Because the company could not cleanly segment access by nationality, the practical result was a full shutoff for every customer. The stated trigger was a national security concern around a jailbreak that let the model read a codebase and spot software flaws. The hosts stress that no one was hacked and no data was breached. Anthropic complied but publicly disagreed, calling it a misunderstanding, and the situation was likened to a car part being recalled mid-deployment.
From there the conversation turns critical (without getting too political). Andrew, who previously worked as a press secretary in Congress, argues that government simply moves too slowly to regulate AI at its current pace, and that pulling a single model is more PR theater than a real solution. Thomas notes the chilling effect: people he knows are now experimenting with DeepSeek and looking harder at European model options, newly aware of how unstable reliance on U.S. models can feel. They also flag the irony that Anthropic's own transparency, admitting no model is perfectly jailbreak-proof, may have handed regulators their rationale. Both agree this sets a dangerous precedent: a government can now take down a frontier model after launch.
The back half is tactical. The core lesson is that single-model dependency is now a real operational risk, on par with relying on one 3PL or one ad platform. Agencies feel it harder than brands because of scale, with one analogy of 70 clients meaning 70 fires to put out. The hosts recommend auditing all AI dependencies (prompts, workflows, MCP configs, Zaps), assigning an internal owner for AI governance, building and actually testing a fallback ladder so models are not hard-coded, keeping a manual or "caveman mode" playbook, maintaining a locked manual database in parallel with AI-driven ones, and understanding the underlying logic of your skills and SOPs rather than trusting bloated AI-written ones. Thomas advocates for open-source systems and local models as insurance. The closing takeaway is that AI is shifting from a playground into pseudo-regulated infrastructure, so contingency planning is no longer optional.
To connect with Andrew Foxwell reach him here Andrew@FoxwellDigital.com
To connect with Will Sartorious DM him here https://x.com/will_sartorius
To Connect mith Thomas Moen DM him here https://x.com/thomasmoen
To learn more about Foxwell Founders and conversations like this one, go here: www.foxwellfounders.com
Full Transcript:
(00:04) If you work in D2C and you use AI and you're wondering what the F is going on every week, this is your podcast, the AI D2C WTF podcast, your home for tactical tips, strategies, and ideas that you can implement right now in your AI workflows to make your brand or agency more money. Welcome to another episode of AI D2C WTF.
(00:26) And today we are talking about Fable, the Claude model being pulled down by the U.S. government for reasons that maybe you know, maybe you don't know, and what happens to your AI stack when this model vanishes overnight. So I'm here with Thomas Moen, my homie from Norway, who is becoming an absolute legend in AI in Europe and in other places too.
(00:53) And just super excited to talk about this because I think it happened and then everyone was like, wait, what's happening with this? And then it just said on Claude, Fable's unavailable. And I think this creates a lot of interesting precedents, right? In terms of like models being pulled down. So, you know, it was available for three days.
(01:09) That's how long Anthropics most powerful model lasted. And it was a great day, Andrew. I had a lot of, I was so happy. Everybody had a lot of memes about it being a summer fling, a hot summer fling, right? Yeah. You know, so look like that's how long it was available for three days before the U.S. government made them shut it off.
(01:31) And if you have anything that's wired into Fable 5 in those three days, right, this episode is your fire drill. We're going to go through what happened, what to do about it, and what are the things you need to check? Because I think there's a lot of pieces that we want to go through. So first of all, you know, what actually happened? Let's talk about this.
(01:48) So on June 12th, actually it was at 521 Eastern. Anthropic received an export control from the U.S. government saying, hey, you need to suspend access to Fable 5 and Mythos 5, two of its most capable models, three days after launch. The order targeted technically access by any foreign national inside or outside of the U.S., including Anthropics' own foreign national employees, because Anthropic couldn't cleanly segment access by nationality and net effect was a full shutoff for all customers.
(02:19) And they did this because there was national security concern around a jailbreak of Fable, basically asking the model to read a code base and spot software flaws. And so, you know, other softwares can already do this, but it was a jailbreak of Fable. But this, you know, is sort of, this could happen to anybody.
(02:38) This could happen to any model. And in fact, it probably already has, you know, and I'm going to go into the final point, which is Anthropic is, so they complied, but they publicly disagreed saying that it was a misunderstanding. So nobody got hacked, nobody had a breach of data, but it was like, recalled like a car part, right? Like, or whatever. So, mid-deployment.
(02:59) Yeah, yeah. And it's still unclear what the jailbreak or the security issues are. And it's been a lot, a lot of back and forth on things. And I think it's also particularly interesting that before the release of Fable 5 and Mythos, Anthropic has been spending a lot of time kind of hyping how dangerous the Mythos model was and kind of how potentially it would break the internet.
(03:26) And then when they finally kind of release it, it only kind of survives three days before it gets shuts down. But no hacking, no breaching of data. And probably it's a mix of something you can recreate on the other models. And I also personally think it has a lot to do with the relationship of Anthropic and the US government as well, which I don't think we need to go into all that kind of the rabbit hole there, but it's definitely a mix with a lot of factors here.
(03:52) Totally. But I mean, we should go into it a little bit. Like, I think it's a couple things to mention. One is, is that I think, you know, being, I used to be a press secretary in Congress in the United States. And one of the things that I noticed by in being in Congress was that there was a lot of people that wanted, that wrote legislation that wanted to actually think about things long term.
(04:12) And then there was the actual members of Congress that wanted things done yesterday. And it wasn't necessarily thoughtful at all around like the approach. Okay. And I think that we now in the current administration in the United States, you know, people can disagree with me politically, but like, it's a technical term would be, I think it's a little bit of an F show of like, what actually is the strategy around AI? A lot of things are an F show, but we're specifically talking about AI today.
(04:36) And, you know, how are we going to go forth on this? And I think because it's rapid pace, it just doesn't match up with how government works. And so they don't know how to legislate or regulate this at all. And so to me, you know, what happened was it, they, you know, some, some idiot who's running the NSA at this point, right in the United States basically goes and says, and they see an issue and they go, Oh my gosh, there's this issue.
(04:59) It's exposing stuff. We didn't know. Oh my gosh, that's big. We should shut it down. I mean, like it was, it was three days. Like they didn't, these models have been able to do this. And for specifically, they like went to Anthropic, I think partially from a PR standpoint to say like, we're reigning in AI, but I don't think it, you know, it really did anything because it's not really solving the core problem.
(05:20) And I think it also shows the dysfunction and the disorganization around how they're going to approach AI as an administration. But I also think it sets like a decent, or I'm sorry, a dangerous precedent that like a government can take down a model, right? Like that's a problem that it can take down a frontier model post launch.
(05:39) And the ramifications, which I'm saying, seeing Andrew kind of from, from, from my side is that a lot of people are now very skeptical in kind of going deep with the American models, right? So I have a lot of friends who've been starting to experiment with DeepSeq. I have a lot of European friends who are kind of trying to understand the European language model, landscape, et cetera, et cetera.
(06:05) So things that you wouldn't really think about a week and a half ago are now on people's minds and people are a bit more aware of how unstable the current situation is in the U.S. Absolutely. I mean, and the irony thing of this, I think is funny is that Anthropics own safety transparency, like area or division publicly admitting that no model is perfectly jailbreak proof.
(06:30) I think that handed the like regulators or the government, the rationale that they needed. Right. So I think it could have been anybody. And I think that that's, that's a big thing to point out. And I, you know, who knows how this is going to continue to evolve because we're at that phase where AI is only continuing to get more powerful as time goes on and it will get exponentially more powerful.
(06:53) And so what is the government going to do to, to attempt to regulate this? And shutting down a model is not like a way to do it. In my opinion, there's going to have to be, you know, a, basically an industry leading group that's going to go through and review these things. And then there probably is going to be a committee in Congress on this, you know, that, I mean, there probably is, but it's not necessarily that effective at this point because they're now it's June, July, August, they're not doing anything in Congress.
(07:16) And, you know, and, and so the, the presidency, they're not the executive branch. They don't know necessarily how to deal with this. And so I think what you see is you see things starting to pop up like the New York law, which came out in June 9th, where you have to put a, you know, disclosure on your ad creative that says, Hey, this is made with AI.
(07:37) There's a similar law that's actually in the European Union that talks about this. And, and the European Union one is, is classically European. It's like insanely broad covers everything. It has no, you know, like regulation arm. It appears the the fines are like, who knows? It's like, yeah, there's, there's no enforcement arm rather.
(07:58) And so it's like, it's a mess. But my, I guess my point is that you're going to have both, you're going to have all this stuff start to percolate in terms of you're not, you're having top-down people pulling models from the executive branch of the United States. You have bottom-up people pulling, you know, like laws in New York state, you know, and it's like all this stuff's going to start to do this.
(08:16) So over the next year, especially, I mean, you're going to have to be ready that you can't be overly dependent upon only one model. Yeah. And this kind of. AI D to C WTF is brought to you by motion. And this episode is brought to you by runneth from motion. And if you run paid social, I mean, obviously, you know, motion, it's a creative analytics platform that everybody uses.
(08:44) But the thing I've been obsessed with lately is runneth. And it's their AI agent that lives right inside of your ad data. The simplest way to explain it, most people use AI by asking it one thing, like, Hey, what did we do yesterday? How'd it go? Runneth is the upgrade of that. You just tag it and ask it to do almost anything.
(09:00) It goes into your account. Does it like build me this week's creative recap, pull the data and write a brief on what to iterate next. Tell me which ads to make next. And it goes on and does this in plain English. So two things that actually make it really special, right? So under the hood, it's, it's Claude, right? And anything Claude can do, runneth can do basically even better as long as you go to access.
(09:21) And second, it's wired from everything motion has. So the AI tags, the creative, the frame by frame breakdowns, and then it's trained by $14 billion in ad spend on motion as customers, right? So it's pretty insane. So the way that, you know, we're using it, and I know some other colleagues are using it, like I looked at what they're doing, and we talked through it.
(09:40) People are, you know, having it drop a morning brief in Slack on how the account's performing. It gets smarter every time it runs. It's less like a tool and more like a creative strategist working while you sleep. So go see it at motionapp.com. That's runneth from motion and make ads that win without getting lucky. You're going to have to be ready that you can't be overly dependent upon only one model.
(10:06) Yeah. And this kind of leads us into why, yeah, this leads us into kind of why, why you, the audience should, should care about this. Why, why do e-commerce advertisers, agencies, D2P, C people, why, why, why should we care about this? But, but because I think, I think we should care deeply about this.
(10:25) And I think it's, it has kind of exposed how vulnerable a lot of people's AI workflows and, and kind of how they look at things and how to kind of build things and what kind of building blocks they're been using that, that suddenly maybe doesn't feel so future-proofed anymore. No, absolutely not. So, you know, why should you care? Let's go through some of the reasons.
(10:46) Number one, single model dependency is now a real risk for you. So if you're building stuff, creative analysis, analysis, competitor analysis, whatever, customer support workflow, anything you're doing in D2C by one specific model, you know, a directive like this can take you down with zero warning. You're not going to have anything backup.
(11:05) So it's, it's the same exposure as if you had one 3PL, you know, or one ad platform, like you have to be cautious and aware of this. I'm not saying, you know, if Fable is, is what you eventually go with, that's the most powerful. I think that's okay. You just have to have a backup, which we'll get into in a minute.
(11:22) I think that's a, that's one. I think number two, agencies feel it harder than brands. You know, if you ran 70 clients through a pipeline of, of static image creation, you know, that's 70 fires you have to deal with basically. So like agencies are going to feel this because the numbers and the scale that you're doing generally are going to be, you know, greater than that of, let's just say one particular brand.
(11:48) So just be aware of where you're tied in. And, you know, having a plan of, as we talked about this in previous episodes, I talked about this with Will about AI governance, who's running AI governance in your organization and what's the workflow as it relates to what are we doing with AI? How are we using it? Where is it currently integrated? Where is it currently integrated into client deliverable and, and client communication and, and client, you know, asset creation, whatever.
(12:16) And what happens if one of these models goes down? Absolutely. What, what, what's the plan? And, and you have to have somebody in your organization that's the captain of that, that really understands and has backup plans for it, because it's going to be really challenging if you don't have. Absolutely.
(12:32) And just to kind of build on the single model dependency we talked about, I also see that a lot of people are locking themselves into codecs or cloud code or something like that. And, you know, it is, I think when you're operating a business, you're very vulnerable if you, so one thing is the models, but it's also kind of the platforms to make sure that your workflows would work just as good with, with codecs or some other large language models.
(13:03) A lot of people we've been talking to, Andrew, are kind of building their own software and are plugging the models into it instead of working kind of in cloud code all day, et cetera. And I think that that's a very smart way of thinking about how to operate a business that makes you less vulnerable when things like this happens.
(13:22) So to kind of be not platform dependent, I think it would be something you should think about going forward. Yeah, totally. I think being not platform dependent and having a fail safe going through. Another thing to mention, right, is if you handle customer order data. So, you know, we call that like the PII stuff, right, which is around customer payments through Stripe or through Shopify or whatever else.
(13:48) One thing that I noticed in doing the research for this episode is there's a 30 day customer retention policy that was shipped with Fable. Yes. And this is their own jailbreak mitigation sort of trade-off. And D to C brands handling that data, that customer data, that changes by model. So you have to be aware of how that changes and what the changes are within each model.
(14:11) And you can just figure that out by asking the model, hey, how are you handling this? But, you know, I think that that is something to be aware of with customer data, right. And how you're, how you're protecting that. And if you were to run into an issue, you know, with, with a customer saying, how's my data being handled or whatever, you have to be able to answer that.
(14:31) So, so that's, I think some of the bigger pieces that we want to talk about why you should care. I'd say that let's get into sort of a little bit more of the tactics around this, which we've gone into briefly, but I think we'd get into more. So, so one audit, the AI dependencies that you have, right. Get your prompts, your workflows, your MCP configs, zaps, whatever, and make a list of who, how they're connected and what is calling what.
(14:59) And like, again, this is sort of that AI governance piece within your organization, have somebody assigned to this or have a team in your, in your agency or your brand assigned to this of how are we doing this? Who's in charge and how are all these pieces talking to each other? So that if one breaks, there's a workflow that it can go on to a varying model and you're not completely stuck.
(15:18) Also, I think it's important to, to build a fallback ladder, you know, so for all the critical workflows, et cetera, make sure you have fallbacks and you have backup models and actually test the backup models. I had a big issue actually a few days ago where my backup model was not working correctly on one of my open clause.
(15:40) So suddenly everything stopped working. So actually testing that the fallbacks are working, I think is very important. Make sure that your models is not hard coded into your workflows, but kind of make it to Opus or Sonnet or GPT or like whatever it is, I think is also very, very smart. Yeah.
(16:02) I think ultimately another thing along with having this like pathway is also, you know, keeping a playbook for manual mode, right? Like if none of this existed, like how, what are the things that you, what are the things that you're going to do to fix this and how are you going to run the business? So having a plan with the team, understanding if we cannot query this, what is the way that we're going through and doing this? You know, we're doing this in our organization now, which is, you know, looking at essentially with a database of all of our customers. We are,
(16:32) we've done it manually. And now we are basically having duplicative efforts where we're updating it manually, but we're also updating it in AI as we continue to build AI. And we're building like a master queryable database within all of our customers, what they've done. But I'm worried that, that, that a model will get pulled down or there's going to be some issue or whatever.
(16:52) And then like, so that's what we're doing. We're basically building two as it goes through so that we have a fallback. I'm not saying you have to do this, but I do think that if we ultimately move into one that's fully AI automated stuff that we're doing with customer database querying and everything, we should at a minimum at a weekly basis, be also updating a manual database in Google sheets or whatever that's locked so that we know, right? Like, I just think if we become overly dependent on this, this is bad.
(17:17) There's already so much talk around people are becoming lazy. They're not letting, they're not doing the work themselves. AI is doing it. You can tell all the posts sound the same. They're all sound homogenous, which I agree with. And I think there's a lot to be said for that because people jump into like making it easier on them.
(17:36) And I mean, LinkedIn comments are an example of this, right? So anyway, so you need to just like be aware and be conscious of that and have a, have a plan. Yeah. And I think building on that, I think you, you need at the minimum to understand the logic of how things work, right? So you have to understand what the AI is actually reading from, what it's going through, the, the, the processes it's doing.
(18:01) You know, we, we, we've gone from SOPs to AI skills basically. And those skills are getting more and more bloated and more and more complicated. I've actually started now going, going through a lot of my core skills and rewriting it from the, the bloat that AI wrote it into something that I can actually understand and, and understand the logic of that has created actually a higher quality of the, of the outcome.
(18:29) But also if you have to go to what you call manual mode or caveman, like how you did it from back in the days, understanding the logic, the workflows, and what the skills that SOPs are actually doing is super, super important. And last, last thing that I'm becoming a bigger and bigger advocate for is also considering more open source systems, right? So I'm a big open claw, for example, guy, which is an open source project you can put into every different large language models out there.
(19:02) So we're trying to build as much as possible outside codecs and outside cloud code, and just let kind of our systems use them when we need it. But we are also now experimenting with having local models running or doing some of the work. So, so again, experimenting with what if open AI or Anthropic went off the map today? How screwed would you be? And what's the plan to make sure that you can actually keep thriving and at least kind of keep surviving in, in a blip like that? Totally. I think it's, I think the big
(19:37) thing to take away from this is that look, AI is moving from, you know, Hey, this is doing cool, cool shit. It does all this too. It's going to continue to move into, to a pseudo regulated infrastructure. Okay. So you have to be ready that government's going to come in and, and it plays a heavy, playing a heavy hand in this.
(19:55) And I think anybody that's going to be in office is going to continue to try to do this because I think the public sentiment around AI is we like it, we use it, but there's also a ton of skepticism in the United States and people also, you know, are, it's easy to be against AI, right? And it's easy to say like, AI, they're too powerful, these companies, whatever, which I don't disagree with necessarily. Right.
(20:16) So you're going to start to see that you're going to start to see more laws like New York, where there's regulated infrastructure that comes up or regulating the infrastructure of disclosures around AI around, you know, whatever. It's like, you need to, we need to be ready. And there needs to be systems that you have in place, knowing that these things could be eliminated quickly. Okay.
(20:35) So again, like it's contingency planning. It's just, it's easy to go all in on this stuff, but you also have to be ready and have a plan for when it's not. So thanks for checking out our quick episode today on this. We thought it was important to drop. If you have thoughts, questions, please feel free to email us or contact us.
(20:54) And we look forward to having you join us. Bye. Bye. Bye.