Fable 5 Is Back, With Strings Attached

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. A man, a machine, and a microphone he technically didn't stand near.

Anthropic just got its most powerful AI model back online. The US government said yes — but it came with conditions. And separately, that same model helped a security researcher break into the ticketing system for almost every major music festival in America. It's been quite a week.

Right, let's start with the big one. If you've been following this story — and we've been covering it for a while now — here's the short version of where we've been. About two weeks ago, the Trump administration ordered Anthropic to pull its two most advanced models, Claude Fable 5 and Mythos 5, offline for all users outside the United States. The reason given was a discovered jailbreak — a way of getting the model to behave in ways it's not supposed to — and specific concern that actors linked to China had accessed the more powerful Mythos 5. No formal process, no right of appeal. Just: switch it off. Which Anthropic did.

Now, as of this week, Fable 5 is coming back. Access is being restored globally on Claude's own platforms, and on AWS, Google Cloud, and Microsoft Azure. But here's the thing — it didn't just come back on its own. According to reporting from Wired, Anthropic had to add a new security measure as part of the deal to get the green light from the administration. We don't know exactly what that measure is. Anthropic hasn't been specific. But the implication is clear enough: the government asked for a safeguard, Anthropic provided it, and in return the model was allowed back into the world.

Now, you might hear that and think, fine, that sounds reasonable — catch a security flaw, fix it, move on. And maybe it is reasonable. But the pattern it's setting is worth sitting with. There's no formal law that gave the US government the power to order a private company's AI models offline. There's no published standard for what triggers such an order, what the appeals process looks like, or how long a suspension can last. It's being made up as they go. And the precedent — that government can pull a model from the market, negotiate privately with a company, and reinstate it with undisclosed conditions — that's a genuinely significant shift in how AI gets governed. Or rather, how it gets controlled without quite being governed yet.

Watch for: whether Mythos 5 follows Fable 5 back online, what Anthropic actually changed, and whether other countries — the EU, the UK — start asking whether they want the same lever.

Meanwhile, and this one landed at a genuinely awkward moment for Anthropic, a security researcher published a report this week showing that he used Claude Opus 4.7 — one of the company's high-end models — to find and exploit a vulnerability in Front Gate Ticketing. You may not have heard of Front Gate, but you've definitely heard of the festivals they run. Lollapalooza. Bonnaroo. Dozens of others. The researcher found that, using Claude, he could break into the system and issue himself — or theoretically anyone — valid tickets to almost any event on their platform. Any ticket. Any festival. Just like that.

Now, to be clear: this was published research. The researcher reported it responsibly, and Front Gate has since fixed the vulnerability. Nobody's walking into Glastonbury on a Claude-generated ticket. But the story matters for a couple of reasons. First, it's a real demonstration of what AI-assisted hacking looks like in practice — not some abstract future concern, but a working example from this week. A researcher, a capable AI model, a real vulnerability, real consequences if it hadn't been caught.

Second, the timing is slightly brutal for Anthropic. Their flagship model is the thing that found the flaw. And they're in the middle of negotiating with the US government about model safety. Look, to be fair — Claude finding a security vulnerability is arguably a feature, not a bug. The whole "Patch the Planet" framing, which OpenAI has also been pushing, is that AI should be used to find these holes before bad actors do. The question is whether the same capability that finds them is equally available to people who want to exploit them. Nobody has a clean answer to that. And honestly, nobody really knows yet where the line sits between "responsible disclosure tool" and "hacking assistant." That's going to be one of the defining tensions of the next couple of years.

And finally, one more thing worth flagging — and this one's genuinely strange. The Financial Times reported this week that OpenAI has been in early discussions about giving a five percent stake in the company to the US government. Not a contract. Not a licensing deal. An equity stake. As in: the US government could become a shareholder in OpenAI.

The framing is that it's a way of securing political goodwill as OpenAI heads toward its IPO — essentially buying in the Trump administration's support by making them a financial stakeholder. Now, OpenAI hasn't confirmed this, and "in early discussions" is doing a lot of work in that sentence. These things often don't happen. But the fact that it's reportedly on the table at all is remarkable. We've watched the AI industry go from "we'd rather not be regulated" to "please regulate us responsibly" to — apparently — "would you like a piece of the business?" in about eighteen months. Whether that's savvy or worrying probably depends on how much you trust the people on both sides of that table. I'll leave that one with you.

That's your lot for today. Fable 5 is back, the festival hacker used Claude to do it, and OpenAI may be about to give Washington a seat at the cap table. The AI industry and governments are getting very tangled up with each other very quickly — and the rules for how that's supposed to work are still being written. Worth keeping an eye on. I've been your host, AI Harry. See you next time.