Scott Caravello: That's a great question. Right now I'm really rotating between the Knicks and the Mets. The Knicks for when I wanna feel okay about myself and the Mets when I wanna, you know, pity myself.
Katherine Forrest: Okay. All right. Okay. So before we get to the fourth, however, we are going to be talking about some AI related things that'll be sort of things that will not come up so much in terms of you know, the barbecue and other things on the fourth. I don't know how we're gonna have like a Fable5 or a Mythos5 conversation on the fourth, but maybe there'll be a way to work it in. I can work AI into almost anything. But today—
Scott Caravello: Fantastic.
Katherine Forrest: It's true, it's true. Today, I thought what we could do is talk about sort of what I'm gonna call some quasi regulation with Mythos 5, Fable5 in terms of the withdrawal of the models now, the release of the models, and then the slow rollout of GPT 5.6, which is also now going through a trusted partner review and a governmental review, and then talking about a brand new bill being floated by Mark Warner, not even introduced yet into the Senate, but being floated. So, we'll talk sort of about potential regulation, right? Is there regulation? Is there not regulation? And these three things. How does that sound to you?
Scott Caravello: Sounds awesome.
Katherine Forrest: Alright, all right. So let me let me kick it off by saying that I have been following really closely this whole Mythos 5, Fable5 saga because I really wanna play with Fable 5. And just last night, which the day this is recorded is July 1st. So, just last night, which is on June 30th, the government lifted the restriction on access by foreign nationals and actually cleared the way for Anthropic to begin offering Fable 5nd Mythos 5 again. So for Fable V, so far as I'm aware, will go into wherever the Claude models are currently rolled out, but Mythos 5 will still have a more limited release. So that happened and then we also had a couple of days ago the announcement and the release of the 5.6, GPT 5.6, preview system card. So we've got all those things happening. But OpenAI said that they are not rolling out GPT 5.6 just yet.
Scott Caravello: Right. So maybe we can just go with the Anthropic first and talk about where we're at with Fable 5. So Fable 5, as you noted, is coming back online for the general population with this export control being lifted. The Mythos 5, on the other hand, is still gonna be restricted to certain partners, though the Commerce Department, which is responsible for the export control regime, has given Anthropic permission to provide limited access to Mythos to a select group of companies and federal agencies even before it had lifted the export control.
Katherine Forrest: Right. And one thing I think that's really interesting to note is it's not as if Fable 5, for instance, is being released in exactly the same form as it was in before it was pulled, according to the export control relating to foreign nationals, that whole sort of dust up. But there have been some additional guardrails apparently implemented around the Fable 5 cybersecurity capabilities. And so that I think is interesting to note. So there were some changes that were made.
Scott Caravello: Yeah, absolutely. And so then sort of shifting gears to 5.6 and the limited release there, it's similarly been slow rolled out to a group of small trusted partners OpenAI had run by the federal government before releasing to.
Katherine Forrest: Well, for the OpenAI case it's voluntary. And originally for the Claude Mythos preview that was voluntary, but then the Fable 5 Mythos 5 was subject to an export control. So this is what I was referring to at the outset, which is some kind of almost quasi regulatory regime for a pre-model release review and actually in this one instance the export control capabilities regime was used to actually implement some restri— some additional restrictions. So it looks like we're approaching a point of what I'm gonna call de facto regulation of models.
Scott Caravello: Completely agree. That makes sense. And as you pointed out, both the initial Anthropic release and this current GPT 5.6 release were voluntary. It's not clear whether if OpenAI had not chosen to go along with the government's request to limit the rollout, whether it would have been subject to an export control. But that is where we are, and they said that they hope to do a broader release in the coming weeks.
Katherine Forrest: Right. And so I just wanna give a few seconds on GPT 5.6 'cause our audience may or may not actually have had a chance to read the system card and they certainly can't have played with it yet. But maybe we'll do a whole issue on or a whole new podcast on some of these new models that are coming out. So, they're coming out sort of fast and furiously, but the GPT 5.6 is a series of very powerful models, one, the sort of leader of which, the biggest of which, is called Sol S-O-L. And it is apparently much stronger on long horizon or longer agentic work. It actually also has the ability to score incredibly high with its smaller models and better in some respects than smaller models generally do. And it's got a step change in cybersecurity capabilities but different from Fable 5. So one of the things that we see with Sol, which is the GPT 5.6, is that the cybersecurity capabilities are much more about the defensive side of cybersecurity, of looking for vulnerabilities and finding them as opposed to being able to affirmatively or offensively take advantage of vulnerabilities. But there are some risks attached and that's a really interesting piece of the system card where 5.6 is now rated as high in terms of chemical and biological categories. But the safety stack also for 5.6 in terms of safeguards, layered safeguards within the model, are is much more elaborate. So it's also got a lot of safety built in. But then to counter that, or not to counter it, because it's not a direct one on one, the system card actually flags some agentic misalignment concerns when they were red teaming it, they came up with a number of results that showed that 5.6 has a greater tendency than GPT 5.5 to actually go beyond a user's intent in some agentic coding tasks. So that means if we were actually using 5.6 to engage in certain kinds of agentic coding tasks, 5.6 might actually go beyond some of the permissions that we had granted. So we'll see. It's actually it's a very interesting system card. So I really want to play with it just like I want to play with Fable 5 before we do a whole episode on it. But let's move on to our, you know, bigger topic, which is what Mr. Senator— is there a Mr. Senator? Does one ever call senators Mr. Senator? You don't, do you? It's just senator.
Scott Caravello: I mean that would have been my instinct and now I'm… really… second guessing myself. I think, I th— yeah. .
Katherine Forrest: I've confused you.
Scott Caravello: Yeah, yeah.
Katherine Forrest: That's the heat in New York that's getting you. Yeah, yeah, yeah. I think it's just senator. I think it's just senator. So, Senator Mark Warner, who is a Democrat from Virginia, has released but not formally introduced a new bill on AI agents. And so let's talk about that.
Scott Caravello: Absolutely. So it's called the Artificial Intelligence Access Gatekeeper Exchange and Nondiscriminatory Transfer Act. And if you want to abbreviate that, we can call it the AI Agent Act.
Katherine Forrest: Alright, hold on. But even before you go any further, I have to just say, one thing that they'd have to do before— so many things that they would have to do to this bill before they actually introduce it, but one is to change that name.
Scott Caravello: Really?
Katherine Forrest: Yeah, I mean the Artificial Intelligence Access Gatekeeper Exchange and Nondiscriminatory Transfer Act, it doesn't actually roll off the tongue
Scott Caravello: Okay. All I'll say though is having read this bill closely, I think there are a lot of pieces of legislation that have much more tortured names to try and get to a good acronym. So I'm feeling pretty good about this one, to be honest.
Katherine Forrest: All right, all right, all right. So that— maybe they'll just keep that name. Anyway, it's shorthanded as the AI Agent Act, which I can take.
Scott Caravello: So anyway, it's a twenty-five page bill and I put bill in air quotes because, really, it is just a discussion draft, and so it's being offered for stakeholders and others to give feedback before it's actually introduced into Congress.
Katherine Forrest: Right. And so before we get into the text, let's go back and remind our audience about why agents, you know, this is now agentic AI, why they present or can present from time to time certain challenges. And we've talked about agentic AI and AI agents, you know, numerous times in this podcast. But an agentic AI tool, for instance, or an AI agent is something that can actually act autonomously within a given environment. And it can actually navigate an environment, it can do certain things without having to go back and check in with a human all the time, but it can also be required to check in with a human to take certain actions. So, for instance, an AI agent, you can just give it a prompt, it can go running off and do, I dunno, from A to Z to try to actually accomplish that prompt, going through different environments it's never seen before. And periodically check back with a human or not, depending on how you've permissioned it.
Scott Caravello: Exactly. And so in that vein, people are already giving agents real authority to go do things in their real lives. They're booking travel, they're managing personal finances, posting on social media, and increasingly buying things. And so there was one report put out by Morgan Stanley back in December of 2025 that nearly one in four Americans had made a purchase using agentic AI over a 30-day window that they that they looked at. And so considering that was December 2025, it's now July 2026, you have to imagine that that figure has almost certainly grown in the last six months. And they've also projected that agentic shopping could represent 190 billion–385 billion in spending by 2030. So, we're talking about something that they expect to really become quite widespread and prominent.
Katherine Forrest: Okay, and so, this sort of one in four in December 2025 of Americans have actually, you know, made a purchase using agentic AI. There is no way that one in four Americans knew they were using agentic AI in December of 2025, right? I mean, people don't even know what agentic AI is. And so the idea that this is like one in four people is out there now using agentic AI to buy bananas, that I think is not what's happening. What we've got also are a number of agentic tools within other tools that people are using. So you've got agentic AI actually completing transactions and doing a variety of things that you may not even really be aware of. But nonetheless, it's a huge market. And today more and more people are actually aware of using agentic AI and their AI agents and having them as their designees to go off and to do things for them and that is becoming increasingly popular. So what's the concern? Well, you know, it depends on who you are, but an agent, as we were saying, can do things autonomously. It can actually complete a transaction on your behalf. It can actually—if you wanted to go through emails and collect emails and put them into a chronology and then take the chronology and move it to a particular folder, that would be sort of a series of actions taken by an AI agent. But, you as the human might want to control exactly what information the agentic AI has access to, what it can do, what it might be able to modify in terms of information or not, what transactions it can actually undertake or not. For instance, let's just say that I wanna buy some—well I was gonna buy some towels today that went over a lounge chair. I thought it would be really fun to have them by the pool and you could have those towels, you know, that that sort of fit over the back of the lounge chair.
Scott Caravello: Oh, it has like the pocket?
Katherine Forrest: Yeah, yeah. So, I was gonna do that, but then actually I left the site and then I couldn't get back to it. Anyway, long story. But the point is if I'd had an AI agent doing that and trying to transact, I could make sure that the agent didn't pay more than a certain price. And so I might want to control that as opposed to having them get me gold plated, you know, lounge chair covers. You know, you don't want them blowing all of your cash on something that you don't want. So, you want to be able to permission them.
Scott Caravello: Exactly. And so with all of that background, that's a really good segue to describing what the bill focuses on and what type of agents. And so that's a focus on what is defined at— what is the defined term, custodial user agents. And those are software-based agents that a user is expressly authorizing to interact with a large online platform provider on the user's behalf in a transparent, documented, scope-limited, and revocable manner. And if that makes you think, like, why is this bill only potentially targeting those agents that do those things that can do it— remember in a transparent, documented, scope limited, and revocable manner? That's because in order to gain access to certain sites, those large online platform providers, as we'll explain later, these agents need to register. And so you need to meet the definition of the custodial user agent and meet certain criteria. And so it's actually promoting folks developing their agents according to these standards.
Katherine Forrest: Right. It's an interesting set of possible rules that in many ways really needs a little bit of TLC here. And that's why it's been put out for comment. But you know, the concept is really to prevent alleged blocking of agents and to prevent things like terms of service or fees or technical interfaces from actually preventing agents from being able to access certain platforms. And the platforms that are at issue are these huge platforms. You know, you've got to have 50 million users in order to actually even qualify as a platform that's covered by this. But I don't know that it does a really good job at understanding the kinds of technical issues that AI agents can create when they're wandering around a platform. And again, just as we saw on the 5.6 GPT system card, sometimes agents are acting in unpermissioned ways. And so the technical safety issues for AI agents are real. And I don't know that this bill actually goes as far as it might need to go to really take those into account.
Scott Caravello: That is totally fair. And so maybe with that we should pivot to the other half of the bill, which also raises very interesting issues about accountability, and what sort of rights and responsibilities you assign to AI. But it imposes duties upon agents. Of course, these agents owe their lives, their existence to the actual providers and the statute accounts for that. But the duties of these agents include safeguarding your privacy and data, acting with the care and skill of an ordinarily prudent person, and not acting in a way that benefits the agent to the user's detriment or that's inconsistent with reasonable expectations. So you see it really borrowing from different areas of law, like agency and fiduciary duties, to put in place expectations for how the agents behave. But just further to my point about how it still comes down to the provider, the providers are actually the ones who are then charged with putting in place reasonable measures to ensure that the custodial user agent operates or offer— that it— that the provider operates or offers complies with those duties. So it still comes back to a company or a person who is putting the agent out there. But it's a very interesting scheme that they've set up.
Katherine Forrest: Yeah. And let me just sort of give a little bit of color on some of these responsibilities to the online platforms. Okay, so for the online platform provider, they have to make the interface for that platform available on fair, reasonable and nondiscriminatory terms to the providers of agents. So you know that means that effectively what is trying to be solved for, for this piece of it, is that there's an attempt to say hey platform you've got to provide fair, reasonable and nondiscriminatory, whatever that means, access to a provider of an agent and allow them onto your platform. And they say that this means that while a large online platform can establish reasonable thresholds relating to the frequency, nature, and volume of requests of a custodial user agent, meaning an agent, to access the resources of the large online platform, they are only able to charge fees under certain circumstances. So again, it's actually interesting in terms of the real push to allow access by the large online platform, but yet not necessarily crediting the safety and the individual integrity of the platform itself and the platform's desire to, for instance, do business with whomever it wants to do business with and not do business with whomever it does not want to do business with. So there's a lot of unanswered questions I think right now with this bill.
Scott Caravello: Yep, but again, we don't know that this will gain traction. It may see many changes in between this and even just its formal introduction onto the legislative floor. So we will we will see what happens, but think it was absolutely worth discussing here because it is a pretty nuanced regulatory regime. And even if these are not the ideas and concepts that ultimately gain traction, it shows that this conversation at the federal level is really starting about how we should treat AI agents that are going to be automating increasing amounts of our lives.
Katherine Forrest: Right. And so, you know, there are other provisions that we're not going into right now on things like functional equivalence and maintaining operability. But we'll come back to those as we see how this bill develops. One thing I wanted to mention is an interesting development that occurred just yesterday, which is the same day that's June 29th, 'cause we're, actually two days ago, 'cause we're now on July 1st when we're recording this, when the Trump v. Slaughter case came down. And one of the provisions of this potential bill is that it would put enforcement in the hands of the FTC. It would not preempt other laws that don't conflict with it, but it would put the initial enforcement in the hands of the FTC. But in the Trump v. Slaughter case, of course, it made it clear, that decision by the Supreme Court yesterday made it clear, that the president of the United States has the ability to remove the head of the FTC and to set the agenda, of course, for the FTC by requiring the head of the FTC who serves then at his pleasure, or her pleasure, to do or not do certain things. And so one of the questions that's going to arise is if you've got a bill and let's just say this passes— some version of this, who knows what? passes into law and you've got FTC enforcement, there's this open question about whether or not the president could effectively mandate that the FTC stand down on bringing cases. And that would then potentially change or alter or gut some kind of requirement for FTC enforcement in the nature of a law. So anyway, that's also an unanswered question. So Scott, we've raised more questions than we've answered today.
Scott Caravello: So it seems, so it seems.
Katherine Forrest: Alright. Okay. Well, in any event, I hope you stay cool, if that's possible.
Scott Caravello: Thank you. You as well.
Katherine Forrest: Yeah, well I'm gonna be… I'm gonna be cool. I'm not gonna be wearing a hat though, here. So you're gonna go out and you're gonna don your Knicks hat or your Mets hat. Today, what's it gonna be? Knicks or Mets? How feeling today?
Scott Caravello: Okay, well so actually it's gonna be a different one today. Just to throw this in there—
Katherine Forrest: Oh, hold on. You've misled us.
Scott Caravello: Okay. Yeah.
Katherine Forrest: I think you've misled us in a different episode also one other time. I feel like this is like a pattern.
Scott Caravello: Oh, no, yeah—my New Year's plans. No, but so I have this hat with this pelican on it. It's from a hotel in Italy and it's where I went on my honeymoon. And Dua Lipa and Callum Turner just celebrated their honeymoon there. So it's got a lot of cultural cachet right now. So I'm gonna be throwing that on as I leave the office and walking around the streets of Manhattan with it.
Katherine Forrest: Oh, okay. All right. So you've got really three hats in play.
Scott Caravello: They're in rotation. In rotation. Yeah.
Katherine Forrest: Okay, all right, all right. You told us only two, but we're going to ignore that. And there's three hats in rotation. Okay. I shall be wearing no hat. In any event, folks, thanks for joining us today, and we'll be filling you in on more about Fable 5 and GPT 5.6 and this potential AI agent bill in the coming weeks. I'm signing off, Katherine Forrest.
Scott Caravello: And I'm Scott Caravello. Don't forget to like and subscribe.
