Anthropic's own testing showed that less capable models reproduced the cybersecurity findings the Commerce Department's June 12 controls were meant to block.
The US Department of Commerce's June 12 export controls on Anthropic's most capable Claude models were framed as a frontier-level intervention: stop the most capable AI systems from helping users find software vulnerabilities and write exploits. Three weeks later, the same controls are coming off, and Anthropic's own technical write-up on the restoration makes the case that the controls were aimed at the wrong tier.
Anthropic said the Commerce Department has lifted the export restrictions on Claude Fable 5 and the more capable Mythos 5, and that Fable 5 will be available globally on July 1 across the Claude Platform, claude.ai, Claude Code, and Claude Cowork, on Pro, Max, Team, and select Enterprise plans. India Today reports that Mythos 5 is returning alongside Fable 5 for users outside the United States. The original directive applied to both models and took effect immediately. Anthropic says it suspended both models entirely because it had no real-time way to verify user nationality and did not want to risk noncompliance.
The original trigger was a report from Amazon researchers who showed that Fable 5's safeguards could be bypassed with a prompt-based attack to identify software vulnerabilities and, in one case, demonstrate exploit code. The Commerce response, signed June 12, treated the finding as a frontier-tier problem. Anthropic's own subsequent testing, summarized in its redeploying-Fable-5 post, complicates that framing. The company says several less-capable models, including its own Haiku 4.5, Sonnet 4.6, and the older Opus 4.6 and 4.7, produced the same vulnerability-identification results. For the single exploit-demonstration case, Anthropic says every model it tested reproduced the result, including Haiku 4.5, Sonnet 4.6, Opus 4.6, 4.7, and 4.8, plus GPT-5.4, GPT-5.5, and Kimi K2.7.
This is the mechanism the Fable 5 return is actually about. If the harmful cybersecurity finding is reproducible on a model from a year ago, the policy assumption that the danger lives at the capability frontier is the part that needs revisiting, not the model that produced the result. Digital Trends reports the Fable 5 restoration. What makes the restoration a regulatory story rather than a product story is Anthropic's choice of the new instrument it is using to ship the model.
That instrument is a set of task-level classifiers that target cybersecurity work rather than a capability-tier cutoff. Anthropic says the new classifiers are designed to block more cybersecurity tasks, and that some routine work, including parts of coding, may be temporarily affected while the system is rolled out. The shift is from "this model class is too capable to export" to "this kind of output is too dangerous to produce from any model without an additional check." The wire framing calls that a jailbreak fix. The policy framing calls it a different theory of the case.
The independent expert reaction has already gone on the record. Katie Moussouris, CEO of Luta Security, who reviewed Anthropic's copy of the Amazon report, called the Commerce response "a complete overreaction," arguing that the disclosed information primarily helps defenders. That is opinion, and it sits in tension with the regulatory instinct to treat frontier disclosures as a national-security problem rather than a defender-asset problem, but it is the framing that Anthropic's own reproducibility finding is now closest to.
There are limits to how far this can be generalized. Anthropic's testing was its own, and the classifier rollout details, including the size of the temporary usage limits and which coding workflows will be affected, are still landing in pieces. The Commerce Department has not, in the materials cited here, given a public technical justification for treating the original Amazon finding as a frontier-tier problem in the first place. The June 12 directive was specific to Fable 5 and Mythos 5, not to a category of model behavior. Mythos 5's return is referenced in secondary coverage and in Anthropic's own tweets, but the primary detail on Mythos-specific restoration mechanics is thinner than for Fable 5.
The thing to watch on July 1 is not whether Fable 5 comes back. It is whether the classifier approach holds for tasks the original Amazon prompt was designed to surface, and whether other frontier-model providers follow Anthropic into task-level filtering rather than asking the Commerce Department to keep redrawing the capability line. If the reproducibility finding stands up, the policy lesson is that AI export controls aimed at a capability tier are catching the wrong variable. The Fable 5 return will then be remembered as the moment that became a documented argument rather than a critique.