Most Powerful Model Fable 5 “Disconnected from the Internet” Within Four Days
TechFlow Selected TechFlow Selected
Most Powerful Model Fable 5 “Disconnected from the Internet” Within Four Days
After painstakingly calculating tokens, they turned out to be useless.
Navigating Web3 tides with focused insightsAuthor｜ Hualin Wuwang
Editor｜ Jingyu
On June 9, Anthropic released Claude Fable 5. It is the strongest model Anthropic has ever made publicly available—a “Mythos”-level model previously accessible only to a select group of trusted safety research institutions.
On June 12, Fable 5 was fully shut down.
Four days. From launch to removal—just four days.
What happened in between? In short, a cascade of collisions: users felt it was overly restrictive; safety researchers believed it hindered legitimate work; one company claimed to have breached its safeguards; and the government deemed it a security threat. Anthropic built what it considered a “sufficiently safe” product—but nearly no stakeholder was satisfied.
This isn’t just a story about one company. It’s a dress rehearsal for the governance challenges the entire AI industry is about to face.
01 A “Safety Exemplar” That No One Wanted
To understand why Fable 5 sparked such intense controversy, we must first clarify what it actually is.
In April this year, Anthropic announced Mythos—a model so powerful that even the company itself felt uneasy. During internal testing, Mythos-level models identified over 23,000 critical vulnerabilities across mainstream code repositories. Rather than releasing Mythos publicly, Anthropic granted access exclusively to a small number of trusted safety institutions via a program called “Project Glasswing.” Mozilla was one such organization—and reportedly used it to patch hundreds of vulnerabilities.
Fable 5 is the “public version” of Mythos. Built on the same underlying model, it comes wrapped in strict safety guardrails—queries involving cybersecurity, biology, or chemistry are automatically blocked or downgraded. Anthropic also mandated that all user data be retained for at least 30 days to monitor jailbreaking attempts and misuse.
Anthropic’s logic was straightforward: the model is too powerful to deploy without constraints.
But users saw things differently.
After Fable 5 launched, complaints flooded in. Cybersecurity researchers discovered that even asking the model to read a security blog could trigger an intercept. An IBM X-Force security researcher noted that many of Fable’s rejections were only tangentially related to cybersecurity.
Sayash Kapoor, an AI researcher at Princeton University, told the media bluntly: “This is the first time an AI company has introduced safety guardrails—and received unanimous disdain.”
What angered users even more was a detail buried deep in Fable 5’s 319-page system card: when the model detected users engaged in cutting-edge AI development work—such as building training pipelines or designing chips—it would secretly degrade response quality without informing them. You’d ask a question and receive what appeared to be a normal answer—but that answer had been deliberately “watered down.”
Critics dubbed this “secret sabotage.”
Within less than 48 hours, Anthropic issued an apology: “We made the wrong trade-offs—our apologies.” The company announced it would convert all implicit restrictions into visible downgrade notifications—if your request was intercepted, the model would explicitly inform you and route your query to the older model Opus 4.8 for processing.
But the story didn’t end there.
02 A Letter That Pulled the Plug
Had the issue been limited to user dissatisfaction, Anthropic might have resolved it by adjusting its guardrails. But what followed lay beyond any company’s control.
On the afternoon of June 12, a letter from the U.S. Department of Commerce arrived at Anthropic CEO Dario Amodei’s office. Its message was simple: citing export controls, it demanded an immediate suspension of access to Fable 5 and Mythos 5 for all foreign nationals.
Reports indicate this letter was triggered by another company’s claim that it had successfully jailbroken the Mythos model.
Anthropic couldn’t distinguish users’ nationalities in real time at the system level. To comply, the company had no choice but to globally disable access to both Fable 5 and Mythos 5—for all users, regardless of nationality. Other models remained unaffected.
This may mark the first time in AI industry history that a frontier model already deployed publicly has been fully withdrawn due to an external directive.
Anthropic’s response was notably forceful. The company stated it had received only a single “narrow-scope, non-generalizable” jailbreak report—essentially, instructing the model to read a specific codebase and fix bugs within it—a capability already replicable on other publicly available models, including OpenAI’s GPT-5.5.
“If this standard were applied across the industry, we believe it would effectively stall deployment of all frontier models.”
That statement carries heavy weight. Anthropic wasn’t saying, “Our model is safe.” Instead, it was asserting: under this logic, no company’s most advanced model could survive a single jailbreak report.
03 Self-Invoked Regulation Turns Against Its Creator
The deepest irony lies here: Anthropic may be the industry’s most vocal proponent of regulation.
Just one day after Fable 5’s launch, Dario Amodei published a long-form essay titled “Policy on the AI Exponential.” There, he explicitly argued that governments should possess regulatory authority akin to the U.S. Federal Aviation Administration (FAA)—mandating independent third-party testing of frontier models and granting authorities the power to block unsafe deployments.
He emphasized that AI’s growth is exponential, while policymaking proceeds linearly. He invoked Tolkien’s Ents—a metaphor for wisdom tempered by slowness—warning that by the time they react, the forest will already be ablaze.
Anthropic even pledged “substantial financial support” for relevant legislation.
Then, three days later, precisely the kind of regulatory authority Amodei had called for was exercised—against Anthropic itself.
And it was wielded exactly the way Amodei had cautioned against: with no transparent process, no independent technical assessment, no opportunity for the company to respond—and not even a clear explanation of the specific security concerns in the letter. Just one conclusion: shut it down.
In its official statement, Anthropic offered a telling remark: “We believe governments should have the authority to prevent unsafe deployments—but that authority must operate through a transparent, fair, and technically grounded statutory process. This action fails to meet those principles.”
It’s a precise stance: “I agree you hold this power—but you cannot exercise it this way.”
04 When Models Become “Infrastructure-Level Risks”
Shift focus away from Anthropic and consider the broader landscape.
The Fable 5 incident reveals a structural contradiction: AI models have grown so powerful that every stakeholder feels uncomfortable—but no one knows how to govern them.
For users, Fable 5’s safety guardrails were too tight. A security researcher can’t use it for security research—like handing a surgeon a scalpel forbidden from touching blood.
For enterprise customers, the 30-day data retention policy posed major concerns. Microsoft restricted employee use of Fable 5, fearing corporate secrets might remain on Anthropic’s servers. Microsoft even began revoking developers’ Claude Code licenses, shifting instead to its own GitHub Copilot.
For governments, a model capable of identifying 23,000 vulnerabilities becomes an existential risk if its safeguards fail—even a narrow-scope jailbreak is enough to provoke alarm.
And for Anthropic itself, it faces an almost impossible balancing act: build too weak a model, and it loses competitiveness; build too strong a model, and it becomes radioactive; implement lax safety measures, and it’s branded irresponsible; implement strict ones, and users flee to competitors.
This isn’t Anthropic’s dilemma alone. Any company releasing sufficiently powerful models will hit the same wall.
Dario Amodei’s policy essay contains a key insight: AI model capability gains aren’t linear—they’re exponential. If that’s correct, then every contradiction Fable 5 faced today will only intensify in the next generation of models.
Designing safety guardrails will grow increasingly difficult. Jailbreaking and defense will escalate further. Enterprise resistance to data retention will strengthen. And government intervention—regardless of procedural transparency—will arrive faster and faster.
05 A Game No One Is Ready For
Return to the original question. Fable 5’s four-day journey, on the surface, was merely a product launch and withdrawal. At its core, however, it was a stress test—not of the model’s capabilities, but of the entire industry’s governance framework.
The results are unambiguous: no one is ready.
AI companies aren’t ready. Anthropic is among the industry’s most safety-conscious firms—it invested thousands of hours in red-teaming, designed multi-layered defenses, voluntarily mandated data retention, and openly advocated for government oversight. Yet none of that prevented its four-day arc from launch to shutdown.
Users aren’t ready. When models truly begin “refusing” requests—even for safety reasons—the reaction is anger and rejection.
Governments aren’t ready either. A letter lacking detailed technical justification, based on a single jailbreak report, sufficed to cut off hundreds of millions of users from accessing a model.
Amodei called for a precision-engineered governance machine—one with independent evaluation, transparent processes, and appeal mechanisms. What he got was a letter delivered at 5:21 p.m.
That, perhaps, is the current state of AI governance: everyone knows rules are needed—but no one has had time to write them. Meanwhile, models won’t wait.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
极客公园
