Last updated: June 2026
Key Takeaways
- On June 12, 2026, the US government ordered Anthropic to block all foreign nationals from its Fable 5 and Mythos 5 models. Because the company cannot filter foreign nationals from its user base in real time, it disabled both models for everyone, worldwide, three days after Fable 5 launched to the public.
- The order rests on a "jailbreak" concern that Anthropic says was demonstrated only verbally and was narrow in scope. Anthropic argues the same capability is available from other publicly released models and that more than 1,000 hours of testing found no broad, universal bypass of Fable 5's safeguards.
- The reason this matters beyond one company: a model hundreds of millions of people depended on was switched off overnight by a third party. That is the clearest demonstration yet of a control property that rented cloud AI has and local, open-weight AI does not.
On Friday, June 12, 2026, Anthropic received a directive from the US government ordering it to suspend access to its two most capable models, Claude Fable 5 and Claude Mythos 5, for any foreign national, whether inside or outside the United States, including the company's own foreign-national employees. Anthropic complied by disabling both models for its entire customer base, since it has no reliable way to separate foreign nationals from everyone else on demand. Every other Claude model stayed online. Fable 5, Anthropic's most powerful publicly available model, had been released only three days earlier.
This article is not an attempt to settle whether the government's security concern is valid. The underlying evidence has not been made public, and we will not pretend to adjudicate a classified national-security claim from the outside. What we can examine is what the episode shows about a question that sits at the center of everything we cover: who actually controls the AI you use, and what happens to you when that control is exercised. For readers who already follow this story, this is a direct continuation of the silent safeguard change we documented two days ago.
What actually happened
Anthropic says the directive arrived at 5:21pm ET on Friday. According to reporting from Axios, the letter came from Commerce Secretary Howard Lutnick and was prepared with officials from the Commerce Department's Bureau of Industry and Security. The instrument used was an export control directive, applied to the models themselves and to the people allowed to use them rather than to any physical product.
A quick definition is useful here, because the term is familiar from a different context. Export controls are government restrictions on transferring specific technology to foreign parties; for years the headline fights were about advanced semiconductors and chip-making equipment. Pointing that same authority at a piece of software already running in production, and at foreign nationals standing inside a US company's own offices, is a notably different application.
The practical chain of events was simple. The order covered foreign nationals broadly. Anthropic cannot identify and gate that group in real time across a global user base. So the only way to comply was to switch the models off for all users. Anthropic's public status page showed Fable 5 as temporarily unavailable on Friday evening, and the company stated plainly that it was complying with a legal directive while disagreeing with the reasoning behind it.
There is one more piece of context that several outlets reported, sourced to an administration official: the government had earlier asked Anthropic to delay the launch of these models, Anthropic declined, and the export control letter followed. We flag that as reported context, not as established motive.
The jailbreak at the center of it
The government's stated concern is that someone found a way to bypass, or "jailbreak," Fable 5's safeguards. Anthropic's account of what it was actually shown is far narrower than that framing suggests. According to the company, the government provided only verbal evidence of a single narrow technique, which amounted to asking the model to read a particular codebase and fix the software flaws in it. Anthropic reviewed a demonstration of the technique being used to surface a small number of previously known, minor vulnerabilities, and states that other publicly available models can find the same issues without any bypass at all.
Anthropic draws a distinction that matters for understanding its position. A universal jailbreak is a method that broadly defeats a model's safeguards across a wide range of capabilities. A non-universal one works only in narrow circumstances. The company says its external bug bounty ran more than 1,000 hours without anyone finding a universal jailbreak, and that no disclosed bypass has produced a harmful result or offered capability that is specific to its Mythos-class models. In its own defense, Anthropic argued that the level of capability in the report it believes prompted the directive is widely available elsewhere, naming OpenAI's GPT-5.5 as a comparison point. That comparison is Anthropic's claim rather than an independently verified finding, and we present it as such.
The government has not publicly detailed its reasoning, and it is entirely possible that officials are acting on information that has not been disclosed. The honest summary is that there are two positions here. The government treats the finding as a national-security risk serious enough to pull the models. Anthropic treats it as a narrow issue that does not justify recalling a product already deployed to hundreds of millions of people. Both of those are real positions, and an outside observer cannot weigh the classified evidence behind the first one.
Why this is bigger than one model
Several outlets covering the suspension reached the same structural observation, and it is the part worth sitting with. This appears to be one of the first times the US has aimed export controls at a deployed commercial AI model rather than at chips or hardware. For years that regime governed the machines AI runs on. Extending it to a live model sets a reference point for how frontier labs ship products, who they are permitted to serve, and where their engineers are allowed to be based.
This also did not happen in a vacuum. On June 2, the administration signed an executive order establishing a framework for the government to review the most capable AI models before public release. That order set up a voluntary process, with agencies getting up to 30 days of early access, and it explicitly declined to create a mandatory licensing or pre-clearance requirement. The documented sequence, then, is a review process built as a request, followed ten days later by the use of a compulsory tool when a company proceeded with a launch the government had wanted delayed. We are describing that sequence, not asserting a plan behind it.
For completeness, this is not the first friction between Anthropic and the current administration. Earlier in 2026 the Defense Department designated the company a "supply chain risk," a label historically reserved for firms tied to foreign adversaries, and Anthropic sued to reverse it; that litigation is ongoing. We mention it as background to a longer relationship, not as evidence of any particular intent in this week's order.
The part that matters for you
Set the politics aside completely. Whether the government was right or wrong, one fact is now demonstrated rather than theoretical: a model that hundreds of millions of people were using was turned off for all of them, overnight, by a party that was neither the user nor, in the immediate sense, the company that built it. Anthropic did not choose to pull Fable 5. Its customers certainly did not. The decision came from outside, and access ended.
This is the principle we return to constantly, made concrete at the largest possible scale: whoever controls the infrastructure controls the experience. When the model you depend on lives on someone else's servers, your access to it is a permission, and permissions can be revoked by parties you never agreed to deal with. The revocation does not require your provider to act in bad faith. In this case Anthropic appears to have fought the order and lost.
It is worth separating this from a related failure mode we covered earlier this week. Two days ago the issue was a safeguard that quietly changed how a model behaved while the model stayed online. This week the model went offline entirely. Those are two faces of the same underlying dependency: when behavior and availability are both controlled remotely, neither is guaranteed to you. We walked through the home-network implications of gated, remotely controlled models back in April, in our look at what Project Glasswing meant for home defenders.
What cloud AI cannot promise, and what local AI actually protects
Here is the honest part, and it cuts against an easy conclusion. Running AI locally does not protect you from the thing the government says it is worried about. If powerful models can help find software vulnerabilities, an open-weight model on your own machine does not make that risk go away; if anything it distributes the capability more widely. Local AI is not a security argument against the government's stated concern, and we are not going to pretend it is.
What local, open-weight AI protects against is precisely the failure mode this week put on display: loss of control over access and behavior. An open-weight model is a file. Once it is on your drive, it cannot be remotely disabled, silently re-tuned, rate-limited, or export-controlled out of your hands. The version you tested is the version you keep. That is not a claim that open models match Fable 5's raw capability; they do not, and on the hardest tasks the gap is real. Sovereignty here is a property of control, not a claim of equal power. We made the same point in our coverage of recent open-weight releases and what they mean for self-hosting.
The table below lays out the trade honestly, on the specific dimensions this episode exposed.
| Dimension | Rented cloud AI (Fable 5 and similar) | Local open-weight AI |
|---|---|---|
| Peak capability today | Highest available; Fable 5 leads on most published benchmarks | Trails the frontier, sometimes by a wide margin on hard tasks |
| Can be disabled remotely | Yes, as demonstrated on June 12 | No; the weights sit on your hardware |
| Subject to export controls | Yes; the model itself can be restricted | Weights already distributed cannot be recalled |
| Behavior can change without notice | Yes; safeguards are configured server-side | No; the file you tested is the file you run |
| Where your prompts go | To the provider; Mythos-class traffic carries 30-day retention | Stay on your own network |
The capability row is real and favors the cloud. The point of running locally is control over access and behavior, not winning a benchmark.
If you depend on rented AI, the practical takeaway
None of this calls for panic or for abandoning cloud models, which remain the most capable tools available and the right choice for plenty of work. The reasonable response is to understand what you are actually buying and to keep a fallback for the things you cannot afford to lose.
A subscription or an API key is a license to use a model, not ownership of it. For casual or experimental use, that is fine. For a workflow that genuinely cannot tolerate a sudden outage or a quiet behavior change, the prudent move is to keep a capable open-weight model running locally as a backstop, even if it is a step behind the frontier. The relevant question is not which model is strongest on a leaderboard; it is which model you can still use next week regardless of decisions made by parties you do not control. For readers ready to set that up, our guide to the best mini PCs for running local AI covers the hardware in practical detail.
If you do keep models locally, fast storage matters more than people expect. Model weights are large, multiple models add up quickly, and load times depend heavily on drive speed. A reliable Gen 4 NVMe drive is the unglamorous component that makes a local-model library pleasant to actually use.
Check Price on Amazon: Samsung 990 EVO Plus 2TB
The broader lesson of this week is not that any one provider is untrustworthy. It is that depending entirely on infrastructure you do not control leaves your access subject to decisions made far above your head. Owning at least part of the stack is how you keep a floor under what you can do.
Frequently Asked Questions
Why did Anthropic shut down Fable 5 for everyone instead of just foreign users?
The government's order applied to all foreign nationals, anywhere in the world. Anthropic has no reliable way to identify and block only that group across a global user base in real time, so the only way to comply with the directive was to disable Fable 5 and Mythos 5 for all customers.
Is Fable 5 coming back?
Anthropic has said it believes the order rests on a misunderstanding and that it is working to restore access as quickly as possible. As of this writing there is no confirmed timeline, and any restoration would depend on the government, not on Anthropic alone.
What is the difference between Fable 5 and Mythos 5?
They are the same underlying model. Fable 5 is the version with safeguards in place for general public use; Mythos 5 is the same model with some of those safeguards lifted, made available only to a small set of vetted security and infrastructure partners. The safeguards are the difference.
Does this affect the regular Claude or Claude Opus that I use?
No. The order applied specifically to Fable 5 and Mythos 5. Anthropic stated that access to all of its other models, including Opus, was unaffected.
Can the government do this to other companies' AI models?
The same export-control authority could in principle be applied to other deployed models. Several observers noted that this case is significant precisely because it points that authority at a live commercial model rather than at hardware, which is why it is being read as a reference point for the wider industry.
Can a local AI model be shut down the same way?
No. An open-weight model stored on your own hardware is a file that no outside party can remotely disable. The important caveat is the software around it: tools that wrap a local model in a cloud service reintroduce the same remote dependency. Keep the runtime local, and the model stays under your control.
