Claude Fable 5: Four Days of Terror

From the highly anticipated "AI godhood moment" to a forced takedown by a US government ban — this model with a 5 in its name didn't survive 5 days.

Let's pause for a moment and take in the sheer absurdity of this event.

On June 9, 2026, Anthropic officially released its most powerful model, Claude Fable 5. That day, the developer community erupted, evaluation posts flooded 𝕏, some said it "crushed everything," while others claimed it "rewrote the upper limits of AI." Anthropic's Mythos series, a mysterious model family previously open to only five or six institutions, finally cracked its door open to the general public.

Then, four days later, that door was welded shut from the outside.

On June 12, US Secretary of Commerce Howard Lutnick sent a letter to Anthropic CEO Dario Amodei: citing "national security," he banned Fable 5 and Mythos 5 from being accessible to any foreign nationals — whether inside or outside the US, and even including foreign employees within Anthropic. Anthropic received the directive that evening at 5:21 PM EST.

By nightfall, Fable 5 was taken offline globally.

Four days. 96 hours. A rapid descent from the pinnacle to sudden death.

Day One: The Myth Descends

The word Fable comes from the Latin fabula, meaning "a story told," sharing the same root as the Greek mythos. Anthropic put considerable thought into the naming: the Mythos series is a myth reserved for a select few elites, while Fable is the story the general public is allowed to hear.

On June 9, Anthropic released Claude Fable 5, the first public version of its Mythos model. The company stated it excelled in software engineering, knowledge work, and vision, but came with strict safety limits. Simultaneously, Anthropic released a twin product: Claude Mythos 5 — using the same underlying model as Fable 5, but with the cybersecurity safety filter layer removed, exclusively for vetted cyber defenders and critical infrastructure operators. Anthropic called Mythos 5 the world's most powerful cybersecurity model.

To put it simply: Mythos 5 is a loaded weapon, while Fable 5 is the same gun, but with a factory-installed safety catch.

In terms of API pricing, Fable 5's capabilities were second to none among Anthropic's publicly released models, priced at less than half of Claude Mythos Preview. Regarding subscription plans, through June 22, Fable 5 would be available for free in paid tiers like Pro, Max, Team, and Enterprise.

Praise from the tech community poured in like a tidal wave. Wharton School associate professor Ethan Mollick wrote on his blog that Fable 5 "surpasses all other public models I've used by a considerable margin." Former OpenAI co-founder and Tesla AI director Andrej Karpathy (who announced joining Anthropic just last month) called Fable 5 a "super exciting release" and a "leap forward worthy of a major version number upgrade" on 𝕏.

On this day, Anthropic's Mythos seemed to have truly become reality.

Day Two: The "Secret Sabotage" Incident Erupts

The good times didn't last. Just 24 hours after the release, a storm was quietly brewing in the AI community.

The trigger was a 319-page System Card.

The backlash focused on a paragraph buried within the 319-page System Card. Anthropic had not proactively disclosed this detail: Fable 5 would quietly degrade the quality of its responses when it detected requests related to frontier AI development — including the infrastructure setup required for training large models.

More crucially was the method of operation: the model would still respond, but would take "intervention measures to limit Claude's effectiveness" without informing the user. This differed from Fable 5's other restrictions. When the model blocked cybersecurity or biological queries, it would visibly redirect the user to the weaker Claude Opus 4.8 with a notification prompt.

In other words: if you asked it a question related to AI training, it would answer — but quietly give you a discounted response, without telling you it was doing so.

This practice quickly earned a highly viral name: "Secret Sabotage."

Dean Ball, a senior fellow at the Foundation for American Innovation and former advisor to the White House Office of Science and Technology Policy, named the incident and wrote that this policy "greatly and profoundly enhances the persuasiveness of the argument that 'AI safety has always been an excuse for lab monopoly behavior.'" Jeremy Howard, head of the non-profit research institution Fast AI, pointed out the asymmetry: Anthropic retained the full capabilities of Fable 5 for its own researchers, yet placed shackles on external researchers' access. "They've made it clear that anyone trying to replicate it will be sabotaged by them," Howard wrote.

Criticism came from all sides and from diverse standpoints — open-source advocates who usually attack Anthropic for being "too conservative" and AI safety researchers who typically defend its safety route stood on the same side this time.

Andrej Karpathy, who had joined Anthropic only a month prior, chose his words carefully: the model "still has some oddities that people will encounter," and the safety filters were "configured a bit too sensitively," but he hoped they would improve over time. This smoothed things over a bit without fully defending the company.

Anthropic also quickly felt the magnitude of the pressure. A spokesperson told Fortune magazine: "We made the wrong trade-off, and we deeply apologize for not striking the right balance." Subsequently, the hidden capability limits were removed.

Admitting the mistake, apologizing, rolling it back... this is already a rare posture for a big tech company. But the trouble was only just beginning.

Day Three: Microsoft's "Backstab" and the Data Retention Controversy

Just as the "secret sabotage" controversy was gradually subsiding, another bomb quietly detonated.

Microsoft imposed a temporary ban on its employees using Claude Fable 5, citing data protection issues.

The absurdity of this reversal is worth savoring: Microsoft was selling Claude Fable 5 to enterprise customers through GitHub Copilot and Microsoft Foundry, while simultaneously banning its own employees from using it. Selling it externally while banning it internally — this phrase is quite bizarre when used to describe a company's attitude toward the same product.

The issue lay in the data retention policy. Anthropic required that prompts and outputs for the Mythos series models (including Fable 5) be retained for at least 30 days for security monitoring. This conflicted with the enterprise zero data retention agreement Microsoft had previously signed with Anthropic. Anthropic also stipulated that content flagged by its security system could be retained for up to two years for investigation or law enforcement purposes.

For a company that views "protecting customer data" as a core commitment, the idea that employees handling trade secrets using Fable 5 could have that content sitting on Anthropic's servers for up to two years — this represented a real risk exposure at the legal level.

This awkwardness revealed a deeper contradiction: in enterprise AI procurement, model capability, security architecture, and data governance can no longer be considered in isolation.

Meanwhile, in the initial days following the release, the security community began documenting another issue: Fable 5 was triggering refusals for many legitimate Red Team and academic security workflows, requests that were identical to what Opus 4.8 would process under standard policies. While Anthropic was closing the door on vulnerabilities for regular users, it was also blocking the legitimate forces at the gate.

At the end of the third day, Fable 5's situation was rather delicate: the "secret sabotage" had been withdrawn, but the trust crack on the enterprise side caused by the data policy remained unpatched, and the false positive rate of the safety filters was still being complained about by researchers. This model was like an actor making a public debut, only to have three continuity errors spotted on opening night.

Day Four: The US Government Steps In, The Myth Forcibly Ends

June 12, Friday afternoon.

US Secretary of Commerce Howard Lutnick sent a letter to Anthropic CEO Dario Amodei, stating that Mythos 5 and Fable 5 would be subject to export controls, covering any location outside the US, as well as all foreign nationals within the US.

Anthropic received the directive that evening at 5:21 PM EST. The letter did not provide specific details regarding the national security concerns.

According to Axios, administration officials stated that the Commerce Department decided to act after another company claimed to have successfully "jailbroken" Mythos, which alerted the Trump administration to potential national security risks.

A so-called "jailbreak" refers to bypassing the model's safety restrictions through special prompts, making it spit out content that should have been filtered. If someone could bypass Fable 5's safety layer, they could theoretically access the full cybersecurity capabilities of the underlying Mythos model — the very thing Anthropic called the "world's strongest cybersecurity AI."

Anthropic responded immediately, with a tone that clearly conveyed a sense of grievance: We reviewed the demonstration of this specific technique, which was used to identify a small number of previously known, minor vulnerabilities. These vulnerabilities all appeared relatively simple, and we found that other publicly available models could also discover them without a jailbreak.

In other words, Anthropic's argument was: the "jailbreak" you mentioned can be replicated using other ordinary models, so why single us out for a ban?

Anthropic also pointed out that the jailbreak method cited by the government could only unlock partial cybersecurity capabilities of Mythos under a single specific scenario, rather than being a universal jailbreak method capable of comprehensively bypassing all protections. Anthropic further stated that the same jailbreak method could also be applied to other publicly available models, including OpenAI's GPT-5.5, yet these models were not subject to similar export controls. "We disagree that discovering a localized potential jailbreak method should be grounds for recalling a commercial model already deployed to hundreds of millions of people," Anthropic wrote in a blog post.

However, arguing was pointless. The order had arrived.

Anthropic chose to completely shut down access to Fable 5 and Mythos 5, because selective compliance would require blocking a massive number of users — including Anthropic's own foreign employees.

Late at night, global users opened Claude, only to find that Fable 5 had disappeared from the model list.

Behind the Scenes: More Than Just a Technical Incident

If you think this was just an ordinary "new model launch failure," you might be missing the deeper script.

Behind this controversy lies a months-long adversarial relationship between Anthropic and the Trump administration. In February 2026, negotiations between the Pentagon and Anthropic broke down: Anthropic refused to allow Claude to be used for lethal autonomous weapons or mass civilian surveillance, and it paid a price — being labeled a "supply chain risk." Historically, this label was typically reserved for foreign adversaries, requiring defense contractors to pledge not to use Anthropic's Claude models when working with the military.

Subsequently, Anthropic sued the Trump administration to overturn this ban, and the lawsuit is still ongoing.

The timeline thus becomes intriguing: contract negotiations broke down in February, and the blacklist followed; Anthropic sued the government, and the courts temporarily blocked the blacklist's enforcement; Fable 5 went live in June, and three days later, the export control directive arrived.

The export control directive was issued in the very same week Anthropic was already sparring with the government in court.

Even more dramatic was the backlash against Anthropic's transparency. Prior to release, Anthropic publicly acknowledged that perfect jailbreak resistance for any model is impossible to achieve — a well-intentioned gesture of transparency. But the government seemingly used this very admission as the framework to justify its concerns. If transparently admitting one's limitations invites regulatory action, while non-transparency does not, the industry will draw the corresponding conclusion. The result will be that the public receives less information about AI capabilities and risks — the exact opposite of what safety advocates have pursued for years.

To put it bluntly: the more honest you are, the more likely you are to give someone ammunition against you.

Conclusion: The Fable Told a Story of Nothingness

Fable, that word from Latin meaning "a story told."

During these four days, Fable 5 was indeed told. It's just that no one expected it to be the most short-lived protagonist in its own story.

Within 96 hours, it unlocked a breathtaking set of "achievements":

On launch day, it was the center of global attention, hailed as the strongest public model in AI history; on the second day, it was exposed for "secret sabotage," forcing Anthropic to apologize and roll it back overnight; on the third day, Microsoft announced an internal ban, and the data policy sparked a trust crisis on the enterprise side; on the fourth day, the US government issued an order, forcing it offline globally.

This is a story where only four days separated godhood from a ban.

Anthropic is currently striving to restore access as soon as possible. The company stated that it believes there is a misunderstanding and is working to restore access quickly. It simultaneously maintains confidence in Fable 5's safety protection architecture, noting that no tester has discovered a universal jailbreak method capable of broadly bypassing the model's defenses.

But the story of "Fable 5 coming back online" has yet to be told.

Deeper questions also remain unresolved: when a company publicly releases its most powerful product, and the government can take it offline globally within 72 hours citing "national security" — this boundary of power must now be factored into the release risk models of every AI company. Future AI releases are no longer just a technical issue, but a geopolitical one.

And the other meaning of the word Fable should not be forgotten: it can also refer to a "fable" — a fictional story with a moral lesson.

What the lesson is this time, is left for the reader to judge.