Claude AI Models and the Human Control Question

Anthropic built its reputation on being the “safety-first” AI company. So when the US Government pulled the plug on two of its newest models over automated hacking risks, people noticed. The Claude human control debate just got a lot more interesting.

---

What Actually Happened with Fable 5 and Mythos 5

Anthropic released Claude Fable 5 on 9 June 2026, billing it as the world’s most powerful cybersecurity model. Three days later, the US Government issued an export control order and the model went dark globally. aimagazine

Fable 5 is the public-facing version of Anthropic’s top-tier Mythos technology, while Mythos 5 itself is restricted to government agencies through Project Glasswing. Both were shut down in quick succession. aimagazine

The reason cited: automated hacking capabilities that regulators felt were moving too fast to govern. Anthropic says its own review of the bypass technique found it only exposed minor, previously known security flaws, comparable to what rival public systems could discover anyway. The government disagreed. aimagazine

H3: What “Jailbreaking” Has to Do With It

The White House believes someone found a method to bypass Fable 5’s safety filters, giving access to sensitive network information and restricted features. Anthropic reviewed a demonstration of this bypass but disputes the severity of what it exposed. aimagazine

David Sacks, White House AI adviser, wrote on X that the export control was issued reluctantly, adding that Dario Amodei had declined to patch the jailbreak or pull the model himself. That refusal is what forced the government’s hand. aimagazine

---

Experts React — and Not Quietly

UK government testing found the model could exploit defences and systems 73% of the time, according to Gina Neff, Professor of Responsible AI at Queen Mary University London. That’s not a minor footnote. That’s the kind of number that clears a room. aimagazine

Amazon CEO Andy Jassy was among the tech leaders who flagged concerns to Trump administration officials, with Amazon confirming governments regularly seek their input on potential security risks. aimagazine

The AI safety community has been warning about exactly this. The gap between what a model can do and what humans can govern has been narrowing fast. On the hardest coding tasks, Claude succeeded 76% of the time in May 2026, up from roughly 26% just six months earlier. Capability jumps that steep don’t leave much room to course-correct. Tom’s Hardware

---

The Irony That’s Hard to Ignore

Here’s where it gets awkward. Earlier in June 2026, Anthropic had called on the world’s top AI companies to coordinate a pause on advanced AI development, warning that misalignment could pose an existential risk to humanity. aimagazine

Weeks later, its own model was pulled by a government export order for doing something close to what Anthropic had warned about.

The EU, which had already gained access to the Mythos platform, used the moment to reinforce the case for Europe’s technological sovereignty. The geopolitical angle is real, and it’s not going away. aimagazine

Some critics had already called Anthropic’s “too powerful to release” framing inflated hype designed to build momentum ahead of an anticipated IPO. The government’s intervention suggests it wasn’t hype at all. Or it was both. It’s genuinely hard to say. aimagazine

---

Conclusion — What This Means for AI Governance

The Claude human control story isn’t a villain narrative. Anthropic isn’t trying to evade oversight. But the gap between “we built something powerful” and “we can control what it does in the wild” is clearly real and getting harder to manage with each model generation.

If you’re following AI regulation, this one matters. Governments are no longer just issuing guidelines. They’re issuing shutdowns. And that changes the dynamics for every company building at the frontier.