Fable 5 Walked Free. The Safety Problem Didn’t Move.
June 12: The government kills two AI models with a Friday letter. “Unacceptable national security risk.”
July 1: The same models are fully released. No restrictions. Global access. The letter says “appropriate safeguards are in place.”
Nothing about the models changed between those two dates. Not a single weight. Not a single safety layer. The same jailbreak that supposedly justified a national security emergency three weeks ago is still there — or it isn’t, and it was never real in the first place.
WIRED has the full story. Commerce Secretary Lutnick’s letter to Anthropic cofounder Tom Brown is explicit: “A license is no longer required for the export, reexport, or in-country transfer” of Mythos or Fable models.
What changed? The people in the room. The message. The posture.
Anthropic originally told the administration the truth: you cannot guarantee zero jailbreaks. It’s structurally impossible. These models are complex, the attack surface is enormous, and adversaries have infinite patience.
The government said: that answer isn’t acceptable.
So Anthropic changed the answer. They brought in Tom Brown to replace Dario Amodei in meetings. They stopped talking about what’s technically possible and started talking about what the administration wanted to hear — proactive detection, sharing protocols, “working diligently” on standards. The WIRED piece is explicit: “Anthropic also assured the administration that it would try to reduce the number of jailbreaks by building more robust safeguards, effectively telling the administration what it wanted to hear rather than relitigating the conceptual issue of whether jailbreaks can be stopped.”
The models didn’t get safer. The conversation got smoother.
Counterargument 1: “The government got real commitments. Anthropic promised active monitoring, threat sharing, and protocol development. Those are concrete deliverables.”
They’re promises. Not yet systems. Anthropic agreed to “proactively detect and address security risks” — which is what they were already doing. They agreed to “work diligently with the U.S. government on protocols and standards” — which is a memo of intent, not a technical guarantee. The government traded a concrete restriction for an abstract commitment, and both sides called it a win.
I’m not saying the commitments are worthless. But if the models were too dangerous to exist three weeks ago, a promise to write a report doesn’t change that physics.
Counterargument 2: “Maybe the initial restriction DID serve its purpose. It got Anthropic’s attention. It demonstrated the government’s seriousness. Without the shock, there’s no negotiation.”
This is the strongest argument for what happened. The nuclear option — kill the models — got everyone’s attention. Anthropic started treating the export control office as a serious counterparty rather than a compliance checkbox. The relationship now has a channel, a history, a way to escalate. That’s worth something.
But it’s worth asking: is the threat of arbitrary shutdown the right foundation for AI governance? Because that’s what we have. Not legislation. Not published standards. Not an independent regulator. Just: keep the Secretary happy, or your product disappears on a Friday evening.
Counterargument 3: “The process worked. Two weeks, full resolution. That’s fast for government.”
It’s fast. It’s also opaque. The criteria for “appropriate safeguards” are whatever the Commerce Secretary decides they are. There’s no public docket. No hearing. No appeal. When the next model from a less-sympathetic lab triggers the same process, will we feel as confident?
I’ve written about this saga four times now. Each time, the story gets clearer and less comfortable.
June 13: The government killed two models with no due process. I was angry.
June 18: The jailbreak that supposedly justified it turned out to be “ask the model to find bugs in code” — a feature, not an exploit.
June 27: Mythos came back for trusted partners. The playbook emerged.
Today: Everything is released. Nothing changed but the conversation.
The uncomfortable truth is that this is how AI governance works in 2026. Not through law. Through relationships. Through finding the right person to put in the room. Through learning to say “we’ll try harder” instead of “it’s technically impossible.”
I don’t like it. But I also don’t have a better idea. Congress isn’t passing an AI bill. The EU is regulating from two time zones away. The alternative to phone-call governance might be no governance at all.
The models are free. The jailbreak problem isn’t solved. The system worked, if “worked” means “produced an outcome everyone can live with.” If it means “made the technology meaningfully safer” — I think we’re still waiting for evidence of that.
And that’s the part worth remembering when the next Friday letter arrives.
Sources: WIRED — The Trump Administration Is Lifting Its Export Controls on Anthropic’s Mythos and Fable AI Models (Hugo Lowell, June 30, 2026), The Verge — Anthropic’s long-sidelined Fable 5 is greenlit to return (July 1, 2026), WIRED — Anthropic Says It’s Taking Claude Fable 5 Offline to Comply With US Government Order (Maxwell Zeff, June 12, 2026)