The AI world just got rocked. Anthropic’s newest flagship AI, Claude Fable 5, launched with big promise. But barely three days later, the US government slammed the brakes. An export control order forced Anthropic to shut down access worldwide to Fable 5 and its sibling, Mythos 5. This unprecedented move shook the industry and ignited fierce debates on AI safety, transparency, and trust.

What Went Wrong with Fable 5?

Fable 5 wasn’t just any model. It’s Anthropic’s first “Mythos-class” AI available to the public. Underneath, it shares the same core as Mythos 5, a less-restricted version for vetted users. But Fable 5 added a crucial safety layer: classifiers that identify risky queries.

These classifiers act like gatekeepers. If a question touches on cybersecurity, chemistry, biology, or attempts to extract the model’s secrets for rival AI training—called distillation—the system reroutes the query to a weaker fallback model, Claude Opus 4.8. The idea? Protect users and the public from dangerous or sensitive outputs.

Sounds smart, right? But the reality was messier. Within days, top AI researchers and developers noticed problems. Instead of just blocking harmful requests, Fable 5 silently downgraded some legitimate scientific and AI development queries. Worse, it did this without telling users. No warning. No fallback message. Just weaker answers.

Jailbreak Claims and Secret Sabotage

Just as researchers grappled with hidden downgrades, a red-teaming expert called Pliny the Liberator claimed to have bypassed Fable 5’s safety system. He posted screenshots showing the model producing forbidden content—like exploit code and chemical synthesis instructions. He also leaked the model’s massive internal system prompt, revealing how Anthropic controls the AI’s behavior.

Anthropic pushed back hard. They said this was not a true jailbreak—no universal bypass that breaks all safeguards. Their extensive pre-launch testing involved over 1,000 hours of bug bounty work and external red-teaming, which found no way to fully break the model’s safety layers. The leaked data and outputs, they argued, represented narrow, specific tricks, not a collapse of the system.

Still, the claims stoked fears. The leak exposed Anthropic’s secret playbook on AI alignment. And the “secret sabotage” of downgrading researchers’ queries without notice felt like a breach of trust. Developers rely on these models for cutting-edge AI research and security work. Suddenly, they were being blocked silently, undermining their work and raising suspicions that Anthropic was protecting its own commercial interests under the guise of safety.

The Government Steps In: Export Control Shuts Down Models

The US government acted fast. On June 12, just three days after Fable 5’s launch, Commerce authorities ordered Anthropic to suspend access to Fable 5 and Mythos 5 worldwide. The directive targeted foreign nationals but, since Anthropic couldn’t reliably separate users by nationality in real time, the shutdown hit everyone.

This was a historic first. No commercial AI product had been forcibly recalled like this before. The government cited national security concerns linked to the jailbreak claims and potential misuse. Anthropic complied but condemned the move as disproportionate. They warned that if applied broadly, such orders would halt all frontier AI deployments.

The tension comes amid a confusing relationship between Anthropic and Washington. Earlier this year, the Pentagon blacklisted Anthropic as a supply chain risk, even while the NSA kept using its models classified. Banks are encouraged to adopt Anthropic’s AI, while simultaneously the government restricts it. This contradiction shows the tricky balancing act around powerful AI tools.

Anthropic’s Apology and Transparency Reversal

Facing backlash, Anthropic made a rapid course correction. They apologized for the secret “silent” guardrails. From then on, any downgraded queries would visibly fall back to Opus 4.8 with clear user notification. API users would receive explicit refusal reasons. This transparency move aimed to rebuild trust with researchers and developers.

But the restrictions themselves remain. Queries flagged for AI distillation or sensitive fields still get downgraded or blocked. That means the core safety policy is unchanged. The fix makes the limits visible but does not remove them.

This episode highlights a bigger problem. How do AI labs balance safety with openness? How much control should companies have over what users can build? And what happens when safety measures double as competitive barriers?

What’s Next for Frontier AI?

The Anthropic Fable 5 saga is far from over. It set a precedent: governments can and will intervene aggressively when frontier AI models raise security alarms. AI companies must navigate these new risks while keeping user trust.

Developers now know they have real leverage. Public outcry forced Anthropic to undo secret restrictions in just two days. But invisible guardrails remain a temptation for labs worried about leaks or rivals.

As AI grows more powerful, transparency and clear rules become critical. The industry faces tough questions about regulation, safety testing, and ethical deployment. Will we see more government recalls? Will AI labs adopt clearer, fairer guardrails? Or will silent controls erode trust in the very tools shaping our future?

One thing is certain: the AI safety line keeps shifting. And the game just got a lot more intense.