The most disruptive month in Anthropic's history is over. On the evening of 30 June the US Commerce Department lifted the export controls that had kept Fable 5 and Mythos 5 offline, and the company began restoring worldwide access on 1 July across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork. The restoration, reported by The Decoder and confirmed by 9to5Mac, ends a suspension that ran roughly nineteen days.
For readers who have followed the saga here, this is the resolution. The ban began on 12 June, when Commerce invoked national security authority to bar access to both models by any foreign national, inside or outside the United States. The trigger was a jailbreak: Amazon's internal researchers found that a particular way of framing a prompt could slip past Fable 5's guardrails, at which point the model identified a software vulnerability and produced code showing how to exploit it. That was enough for the government to pull the plug on a frontier model mid-deployment, a first for the industry.
A new classifier, and a familiar tradeoff
Anthropic did not simply flip the switch back on. It trained a new safety classifier aimed squarely at the prompt-framing technique Amazon uncovered, and the company says it now blocks that technique in more than 99 percent of cases. The catch is the one every content filter runs into. A stricter net catches more of what it should and more of what it should not, and Anthropic acknowledges the new classifier throws more false positives, flagging harmless programming and debugging requests that happen to resemble the blocked pattern. Developers who lean on Claude for security research may find it more cautious than before.
The company is candid about the limits of the fix. It says it is "probably impossible to make any AI model fully robust to jailbreaks," and rather than claim otherwise it is working with Amazon, Microsoft, and Google to set shared industry standards for handling this class of vulnerability. That is a more honest posture than the usual reassurance, and it fits a month in which the gap between a model's safety on paper and its safety under a determined attacker became a national security question.
The terms of the return
Access comes back with a throttle at first. Through 7 July, Pro, Max, Team, and select Enterprise plans can spend up to half of their weekly usage limit on Fable 5 before usage credits kick in. Restoration on the big cloud platforms, AWS, Google Cloud, and Microsoft, is still in progress. Mythos 5, the less restricted sibling, stays limited to approved US organisations under the Glasswing programme, which received government sign-off on 26 June.
The episode leaves a mark that a switch cannot undo. For nineteen days the world's most capable widely available coding model was a lever the US government chose to pull, and the reason was a safety flaw a rival lab found first. The models are back. The precedent that a jailbreak can take one offline is back with them.
Sources
- i. the-decoder.com
- ii. 9to5mac.com
- iii. www.techtimes.com
- iv. cryptobriefing.com
Commentarii · 0