Date Published
June 19, 2026
Total Read
6 min
Tags
On 12 June 2026, Anthropic disabled access to two of its most capable models, Claude Fable 5 and Claude Mythos 5, for all users worldwide. Not for foreign nationals only. Not for one region. For everyone, including paying enterprise customers and Anthropic's own staff.
The reason was an order from the US government. It directed Anthropic to block access to both models for foreign nationals, whether inside or outside the United States, including foreign employees working at Anthropic itself. There was no practical way to comply with that order while keeping the models running for anyone, so the company switched both off completely.
Anthropic's other models, including Claude Opus, were not affected.
The public reporting paints a chain of events that began before most people had heard about any of this.
According to The Information, the first move came from inside the tech industry rather than from the government. Tech leaders, including Amazon's CEO Andy Jassy, reportedly raised concerns with US officials about a jailbreak found in Mythos 5. That detail is worth sitting with for a moment. Amazon is one of Anthropic's largest investors and infrastructure partners, so this was not a rival trying to slow down a competitor.
Those concerns reportedly reached the US National Cyber Director, who called a meeting with senior White House officials. Export controls were chosen as the fastest way to restrict the model, since the only realistic way for Anthropic to comply was to take it down for all users at once.
This is where the story splits into two readings, and both are worth knowing.
The less troubling reading is that an inexperienced new cyber director, under pressure from business leaders worried about AI's risk to critical infrastructure, acted quickly and somewhat bluntly after hearing there was a working jailbreak. On this reading, the response was an overcorrection rather than a targeted move against Anthropic.
The more troubling reading starts with Anthropic's own account. The company has said the vulnerability behind the jailbreak was narrow and similar to weaknesses already known to exist in other frontier models, including recent releases from competitors. Anthropic also points to its own published safety testing, in which Mythos and Fable score far ahead of other frontier models in resistance to prompt injection, a technique in which a model is tricked into ignoring its own guidelines. If that benchmark holds, it raises a fair question about why this particular model was pulled.
There are also conflicting accounts of what the jailbreak actually demonstrated. Some independent security reviewers who looked at the case reportedly concluded that what was flagged looked like ordinary vulnerability patching work, the kind of activity a defender would do, rather than evidence of an attack capability.
Anthropic has been direct about its disagreement with the decision. The company has said it accepts that governments should be able to block unsafe deployments, but only through a process that is transparent, fair, and grounded in technical facts. It says this action did not meet that bar, and it has called the whole episode a misunderstanding it hopes to resolve quickly.
This did not happen in isolation. Anthropic and the current US administration have clashed before. Earlier this year, the Department of Defense revoked Anthropic's access to one of its facilities after talks between the two broke down, and department officials have continued to criticize the company publicly since then.
There is also a political dimension that some commentators have pointed to. Other AI labs have invested heavily in lobbying through political action committees, while Anthropic has been notably absent from such activity. Whether that difference in approach has any bearing on how each company gets treated by regulators is impossible to prove from the outside, but it is a fair question to ask when one company faces sudden export controls, and others do not.
None of this proves intent either way. It is entirely possible this is exactly what it looks like on the surface, a rushed decision following a genuine jailbreak, made by officials who are themselves new to the scale of the risks involved in frontier AI. It is also possible that the underlying motives are more political than technical. Both readings have circulated, and a single jailbreak report involving a model used by hundreds of millions of people is a strange trigger for such a drastic response, regardless of which reading turns out to be correct.
I am not writing this because of which model got blocked. I am writing it because of what it reveals about how fragile access to frontier AI can be.
A model that hundreds of millions of people relied on disappeared within hours for reasons unrelated to its users. No warning beyond a short window to comply. No say in the matter for the businesses or individuals who had built workflows around it.
For anyone working in design, product, or engineering who has relied on a single AI provider for daily work, this is a useful prompt for thinking about resilience.
A few practical things worth doing:
Map which of your workflows depend on one AI provider, and be honest about how much would break if that access disappeared tomorrow.
Build fallback options into critical processes where feasible, even if it adds friction day-to-day.
Pay attention to the policy environment around frontier AI. It is moving fast, the rules are still being written in real time, and the effects of a single decision can land immediately and globally.
Jailbreaking remains an unsolved problem across the entire industry. No provider, large or small, has produced a model that resists every attempt to bypass its guidelines. That is worth remembering the next time a single jailbreak report is used to justify taking a major capability away from everyone who relies on it.
I will be watching to see whether access to Fable 5 and Mythos 5 returns, and how quickly. The answer will say a great deal about whether this was a misunderstanding or the start of a new pattern in how frontier AI gets governed.
Read more: https://www.anthropic.com/news/fable-mythos-access
