Anthropic restores its flagship AI after introducing safeguards that satisfied US regulators.
Most AI launches follow a familiar pattern: a company unveils a faster model, shares benchmark results, and quickly shifts focus to the next release.
Claude Fable 5 did not get that luxury.
Just three days after Anthropic launched the model, it was taken offline—not due to a technical failure or server overload, but after US authorities introduced emergency export controls amid concerns about potential cybersecurity misuse.
Now, following weeks of negotiations and the introduction of new safety safeguards, Fable 5 is being restored for users.
For Anthropic, the redeployment marks the end of an unusual chapter, while for the AI industry it could signal the start of a new one.
The company says Claude Fable 5 is once again available via Claude.ai, Claude Code, Claude Cowork, and its API after Washington lifted the restrictions on June 30. Access through cloud providers is being reinstated separately.
The underlying model has not fundamentally changed.
What has changed is the protective layer built around it.
Anthropic says it has introduced a new classifier designed to detect prompts aimed at uncovering software vulnerabilities or generating exploit code. When such requests are identified, they are redirected to the older Claude Opus 4.8 model instead. The company says the safeguard blocks the reported bypass technique in more than 99% of tests, while also acknowledging that some legitimate coding queries may be incorrectly flagged.
The episode offers a glimpse into where frontier AI is heading.
When Fable 5 launched in early June, Anthropic described it as the first publicly available Mythos-class model—a system designed for advanced software engineering, long-running reasoning, and autonomous knowledge work, while depending on external safeguards to reduce misuse. Its restricted counterpart, Claude Mythos 5, remains accessible only to approved organisations through Project Glasswing.
What drew attention, however, was not its benchmark performance, but the government’s response.
According to Anthropic, the export controls were triggered after a report from Amazon researchers identified a method to bypass one of Fable 5’s cybersecurity safeguards. The company says its own testing showed that several other advanced AI systems could produce similar outputs, but it still collaborated with US officials to introduce additional protections before the restrictions were lifted.
The incident has fuelled a broader debate within the AI industry. Some developers argue that regulating individual models may not be sustainable when competing systems already offer similar capabilities, while cybersecurity experts maintain that tighter oversight is becoming increasingly necessary as AI systems grow more powerful.
Anthropic has also launched a HackerOne bug bounty programme to incentivise researchers to report jailbreaks and other vulnerabilities before they can be exploited. The company says it will continue working with government agencies as it develops future frontier models for release.
