Anthropic says it now stops that technique in more than 99% of tries, as of the June 30 write-up.
The U.S. export-control move: June 12 order, June 30 reversal
On June 12 the U.S. Commerce Department ordered export controls that required Anthropic to cut off two models—Claude Fable 5 and Mythos 5—to any foreign national, inside or outside the United States. The rule took effect immediately; Anthropic, unable to verify every user's nationality in real time, shut both models down for all customers. On June 30 the Commerce Department lifted those controls, and Anthropic restored Fable 5 to users worldwide on July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork.
What triggered the emergency action: an Amazon-reported jailbreak
The trigger was a jailbreak discovered by Amazon researchers. By Anthropic’s account, the prompt used in the jailbreak caused Fable 5 to flag a few software flaws and, in one case, to write code showing how a flaw could be abused. Anthropic characterized the finding as routine defensive security work and said the same requests work on “plenty of weaker models,” including its own Claude Opus 4.8, OpenAI's GPT-5.5, and China's Kimi K2.7. The government and Amazon — the partner that reported the jailbreak — judged the behavior serious enough to justify the emergency controls.
Anthropic’s technical fix: a classifier, fallbacks, and trade-offs
Anthropic responded by training a safety filter, called a classifier, to detect the exact technique described in the report and block it. The company reports the classifier blocks that technique in more than 99% of attempts as of its June 30 write-up. When the classifier blocks a request, Anthropic routes the interaction to the weaker Opus 4.8 model and informs the user. Anthropic acknowledges the trade-off: the new filter produces more false alarms on routine coding and debugging even as it reduces the specific jailbreak risk.
Mythos 5, government coordination, and industry scoring proposal
Mythos 5, which shares the same underlying model as Fable 5 but has fewer safety guardrails, remained under tighter controls. Anthropic restored limited access to Mythos 5 on June 26 for roughly 100 U.S. companies and federal agencies that defend critical infrastructure and said it is still working with the government to widen access. Commerce Secretary Howard Lutnick, who signed off on the reversal, said his department had spent two weeks reviewing the models with Anthropic; in exchange the company agreed to hunt for security problems, coordinate on future launches, and report malicious use it observes.
Anthropic also proposed a shared rubric with partners including Amazon, Microsoft, and Google to score jailbreaks on four dimensions: capability gain, breadth, ease of weaponization, and discoverability. For the most severe cases—Anthropic gives the example of jailbreaks that enable attacks on power grids or banks—the company said it would deploy fixes immediately once severity is confirmed and has stood up a team to monitor jailbreak reports around the clock.
HackerOne, preview access, and the wider regulatory context
Anthropic opened a HackerOne program to let researchers report new Fable 5 jailbreaks and promised the U.S. government earlier access to test future frontier models before release. The episode unfolded against the background of a June 2 executive order that created a voluntary path for companies to have frontier models reviewed and established a classified benchmark to decide which models count as "covered"; the source notes that Fable 5 never went through that voluntary path and that the government resorted to export controls when it wanted to move quickly.
How technologists, policymakers, and defenders are likely to respond
- Technologists and security teams: will watch Anthropic’s classifier, the HackerOne program, and the rubric for rating jailbreak severity; they will be attentive to the classifier’s reported >99% blocking rate and to the increased false positives on legitimate debugging tasks.
- Policymakers and regulators: now have an operational precedent for using export controls when the voluntary review path is not used; the Commerce Department’s two-week review and the conditions Anthropic accepted illustrate a negotiated, case-by-case approach rather than a mandatory licensing process.
- Defenders and procurement leaders in critical infrastructure: have already seen constrained access to Mythos 5 restored for about 100 U.S. companies and federal agencies and have received Anthropic’s promise of earlier government testing for future frontier models.
The immediate crisis is over. The bigger question is not: Washington has shown that when it needs to move fast on a frontier model it can improvise oversight, but the episode underlines that there remains no binding, standardized process for handling such launches.
