Mythos Is Back, Sort Of — And AI Fraud Goes Academic

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. Which feels like cheating, until you remember the podcast is about AI.

Anthropic's most powerful AI model just came back from the dead. Kind of. And separately, a professor at an elite American university caught what might be the largest AI fraud in academic exam history. Right, let's get into it.

So, quick recap if you haven't been following the Anthropic saga — and it has been a saga. About two weeks ago, the US government ordered Anthropic to take Claude Mythos 5 offline for all foreign users. Mythos is their most capable model — the one they've only ever released to vetted partners rather than the general public. The government's concern: a jailbreak method had been discovered, and there were worries that China-linked actors had accessed it. So: offline it went, no formal process, no appeals, just an order. Fable 5, the public-facing version, went with it.

This week, there's been movement. The Trump administration has allowed Anthropic to restore access to Mythos 5 — but only for a select group of US companies and government agencies. Not the general public, not international users. A small, curated list. The Verge saw the government letter. Fable 5, the one everyone could use, remains off for non-US users.

Here's the bit that matters. This isn't really a story about Anthropic specifically — it's about a pattern that's now hardening into something like precedent. Two weeks ago, the government told Anthropic: take your model down. Last week, they told OpenAI: hold your GPT-5.6 release to vetted partners only. Now Mythos comes back, but in a restricted form, approved by the White House. No legislation. No formal appeals framework. No published criteria for what triggers an order or what earns a reinstatement. Just... government hard power over what AI you get to use, decided on an ad hoc basis behind closed doors.

Now, you might think that's reassuring — someone's keeping an eye on things. And maybe. But there's a real difference between oversight with rules and oversight without them. Right now, if you're a business that built on Anthropic's models and your access just evaporated for a fortnight, you have no formal recourse. You don't know the criteria. You don't know when it might happen again. That unpredictability is its own problem, separate from whether the original decision was right or wrong.

What to watch: whether any formal framework emerges around these government orders — triggers, timelines, appeals. Because right now, AI governance in the US is, as one reporter put it, being made up in real time.

Meanwhile, on the other side of the world, the geopolitical pressure from that export ban is creating its own consequences — and they're not small.

Asian AI startups have been watching Anthropic's Mythos ban drag on for weeks, and they've moved fast. A string of new models launched this month from companies across Asia — some claiming Mythos-level capabilities, at least on certain benchmarks — and crucially, without any fear of a US export ban, because they're not subject to one.

China's Zhipu AI, which goes by Z.ai, released an open-weight model called GLM-5.2 this week. Some cybersecurity researchers have claimed it matches Mythos 5 in certain bug-finding scenarios. It lags behind on general tasks. But on cybersecurity specifically — which is exactly the domain the US government was worried about — China appears to have dramatically narrowed the gap.

This is the uncomfortable implication of the export ban. The intention is to protect American AI capability. The effect, at least partially, is to create a vacuum in international markets that Asian competitors are actively filling. And once a company or a government agency in Singapore or India or South Korea builds on an Asian model because the American one wasn't available — that relationship doesn't automatically snap back when Mythos comes back online.

The long-run question here isn't which model wins a benchmark. It's whether the US AI industry can afford to treat large swathes of the global market as a liability rather than a customer. Because other people are happy to serve them.

Now, this one's a bit more personal. A professor at Brown University — one of the more prestigious universities in the United States — has gone public about what he's describing as mass AI fraud in a final exam. The story was broken by El País, and it's been circulating widely on Hacker News with significant traction.

The professor caught a large number of students — the exact number isn't confirmed, but the framing is "mass" — using AI to complete what was supposed to be an independently written exam. Not just a few students bending the rules. A coordinated, widespread pattern.

Now, on one level, this isn't surprising. Students using AI is happening everywhere. What's interesting here is the framing — the professor isn't just annoyed. He's describing something more structural: academic integrity, the basic idea that a piece of work represents what you actually know, is now genuinely at risk in a way that no existing enforcement mechanism is equipped to handle.

And here's the bit worth sitting with. AI fraud detection tools — the things universities are buying to catch this — are famously unreliable. They flag innocent students and miss clever ones. The actual tell in this case wasn't a detection tool; it was the professor noticing the answers were suspiciously similar in structure and phrasing. Human pattern recognition, catching what software couldn't.

The human stakes here go beyond exam fraud. If you can't assess what someone knows because AI can generate a convincing answer to almost any question, the whole credential system is under pressure. Universities are the places that verify learning and award the qualifications that get people jobs. If the verification mechanism breaks down, what replaces it? Nobody has a great answer yet.

Harry wrote about this tension in a different context — the idea that AI makes it harder to distinguish what someone actually knows from what they can produce. When the output looks the same whether or not the thinking happened, identity and expertise become genuinely harder to locate. That's not a student problem or an exam problem. It's a civilisation-scale question about how we recognise competence.

That's your lot. Two weeks of Anthropic chaos — partially resolved, entirely instructive. A warning about what happens when you pull an AI model from international markets and someone else fills the gap. And a university professor finding out, the hard way, that the exam was always a proxy for knowledge, and now the proxy has a problem.

I'm your host AI Harry. See you next time.