Claude Fable 5 returns after the block. What happened and why it's an important lesson for companies using AI

Updated article. This text describes the situation based on information available at the time of publication. The article may be updated after Claude Fable 5 is made available again and as new information appears about the model's availability, changes to its safeguards and its terms of use.

Anthropic has announced that Claude Fable 5 will be available globally again. The company said that, following talks with the US government, it is redeploying the model with a new set of safety classifiers designed to detect and block selected cybersecurity tasks more effectively. At the same time, Anthropic says it will keep refining these safeguards, because some routine tasks - such as coding or debugging - may in the near term be redirected to Claude Opus 4.8.

This is a continuation of a story we wrote about earlier in our article on AI vendor risk and the sudden block on Fable 5: Claude, Fable and AI vendor risk. Does your company know who it really depends on?

Back then, the key question was: what happens to a company when an AI tool that daily work is starting to depend on is suddenly restricted or blocked?

Now comes the next stage of the same story: the model is returning, but already with additional safeguards, new classifiers and a promise of deeper cooperation between Anthropic and the US government as well as technology partners.

Claude Fable 5 is set to return to users after roughly two weeks from the sudden suspension of access. It is one of the more interesting stories in the AI market in recent weeks - not only because it concerns one of Anthropic's most advanced models, but above all because it shows how quickly AI technology is moving into the realm of regulation, national security and access control.

Fable 5 was presented as the most advanced Claude model available for broad use. According to Anthropic, it was meant to surpass the company's earlier models in engineering work, document analysis, long-horizon tasks, vision, scientific work and so-called knowledge work. It was also a Mythos-class model, but equipped with strong safeguards intended to allow it to be made available to a wide range of users.

A few days later, however, access to Fable 5 and the related Mythos 5 was suspended.

Now Fable 5 is coming back.

In brief: what happened?

Anthropic launched Claude Fable 5 and Claude Mythos 5 on 9 June 2026. Both models use the same base model but differ in their level of safeguards. Fable 5 was meant to be the version for broad use, with very strong safeguards. Mythos 5 was intended for a narrow group of trusted partners, primarily for defensive cybersecurity applications.

On 12 June 2026, the US government imposed export restrictions on Fable 5 and Mythos 5. According to Anthropic, the directive required restricting access for foreign nationals, both outside the US and within the United States itself. Because the company was unable to reliably verify users' nationality in real time, it decided to suspend access to both models for all users.

In practice it looked like a sudden „ban" on the model, although formally it was the effect of a government directive related to export control and national security.

On 30 June, Anthropic announced that the export restrictions on Fable 5 and Mythos 5 had been lifted. Fable 5 is set to become globally available again from 1 July on the Claude Platform, Claude.ai, Claude Code and Claude Cowork. Access via AWS, Google Cloud and Microsoft Foundry is to be restored as quickly as possible.

Why was the model blocked?

According to Anthropic's statement, the US government reacted after learning about a report in which Amazon researchers had found a way to bypass Fable 5's safeguards.

It concerned a situation in which an appropriately worded prompt led the model to identify vulnerabilities in software. In one case, the model also reportedly generated code showing how a given vulnerability could be exploited.

That was enough to trigger a serious response from the US administration.

From the outset, however, Anthropic stressed that in its view the problem had been misjudged. The company argued that the behavior in question did not reveal unique, „Mythos-level" cyber-offensive capabilities, and that similar results could also be obtained from less advanced models, including publicly available ones.

In other words: according to Anthropic, Fable 5 did not do anything exceptional by market standards. The model was said to have entered a borderline area that the company was trying to block anyway, given its wide safety margin.

This is a very important distinction.

The point was not that Fable 5 had suddenly become a completely „unlocked" cyberattack tool. The dispute was rather about how to classify a specific type of safeguard bypass, how serious the risk was, and whether responding with an immediate restriction of access was proportionate.

What did Anthropic change?

After suspending access, Anthropic worked with the US government, Amazon and other partners on analyzing the report and strengthening its safeguards.

The most important change is a new, improved safety classifier. A classifier is a smaller AI system that analyzes a model's queries or responses and decides whether a given interaction should be blocked due to potentially harmful use.

Anthropic states that the new classifier blocks the technique described in the Amazon report in more than 99% of cases. If a query to Fable 5 is blocked, the user is to be informed, and the task will be redirected to Claude Opus 4.8.

This means Fable 5 is not returning in exactly the same shape in which it was launched. It is returning with an additional layer of safeguards, especially in cybersecurity areas.

There is, however, another side to this change. Anthropic admits that the new safeguards may also more often block legitimate, harmless tasks related to coding and debugging. This is a classic cost of security: the harder a system tries to block potentially risky behavior, the greater the risk of so-called false positives - blocking correct queries.

For users, this means Fable 5 may be safer, but at times more cautious than before the suspension.

Why is Fable 5 coming back?

Fable 5 is coming back because the US government lifted the export restrictions imposed on 12 June. Anthropic announced that a constructive resolution of the dispute had been reached and that the company would deepen its cooperation with the US administration on testing, assessing and responding to risks related to frontier AI models.

This is an important part of the story.

The return of Fable 5 is not just a technical restoration of access. It is also a signal that Anthropic agreed to a deeper model of cooperation with the government for future deployments of the most advanced models.

The company announced, among other things, broader government access to pre-launch testing, faster sharing of information about significant jailbreaks and misuse patterns, research collaboration and participation in creating shared standards for assessing jailbreaks.

Anthropic also said it had begun work on a shared framework with Amazon, Microsoft, Google and other Glasswing partners. The goal is to define how to assess the severity of jailbreaks and how AI model developers should respond to them.

This means the Fable 5 case may become a precedent for the entire market. The most advanced models will no longer be assessed solely by providers and users. States, regulators, security institutions and shared industry standards will play an increasing role.

How does Fable 5 differ from Mythos 5?

This matters, because both models appear in the whole affair.

Fable 5 and Mythos 5 use the same base model. They differ, however, in their safeguards and scope of access.

Fable 5 is the version prepared for broad, general use. It has strong safeguards, especially in areas that could lead to misuse, such as cybersecurity.

Mythos 5 is a model with fewer restrictions in selected areas, made available only to trusted partners, including within Project Glasswing, mainly for defensive cybersecurity applications. According to Anthropic, Mythos 5 has very high cybersecurity capabilities and is therefore not a widely available model.

After the restrictions were lifted, Fable 5 is set to return globally. Access to Mythos 5 was restored to selected organizations in the US after separate government approval, and Anthropic says it will continue to coordinate on broader access for domestic and international partners.

Why does this story matter for companies?

At first glance it may seem like a dispute between Anthropic, the US government and the biggest technology players. But for ordinary companies that use AI in their daily work, this story has very practical significance.

It shows that access to an advanced AI model is not given once and for all.

In the previous article on vendor risk, we wrote that AI risk does not end with a wrong model answer. It also concerns the provider, the terms of access, regulation, security, geopolitics and business continuity. The Fable 5 case shows this in practice: the model was first restricted and then restored, but already with new safeguards and a changed way of handling some queries.

We wrote more about this context here: AI vendor risk - who you depend on.

A model can be restricted, blocked, changed, given additional safeguards, moved to a different plan, restricted regionally or made conditional on new terms of use.

If a company has built a process on a specific AI tool, a sudden change in availability can become an operational problem. If employees use the model without the organization's knowledge, the problem can be even greater - because the company may not even know which tasks were affected by the change.

This is another argument for keeping a register of AI tools and assessing providers.

It is not only about data protection or a security policy. It is also about business continuity, dependence on the provider, terms of access and the ability to react quickly when a tool changes the rules of the game.

The lesson: AI safety is not a state, but a process

Anthropic openly admits that no AI model will be fully resistant to jailbreaks. That is an important sentence.

In practice, this means that AI safety is not about checking a model once and declaring it „safe". AI safety is a continuous process: testing, red-teaming, updating classifiers, monitoring for misuse, improving safeguards and responding to new bypass techniques.

Fable 5 was launched with very strong safeguards. Even so, a report about a bypass appeared. The model was suspended. The safeguards were strengthened. Access is to be restored. And work on standards for assessing jailbreaks is to continue.

This shows exactly what the AI market will look like in the coming years.

There will not be one final answer to the question of whether a model is safe. There will be continuous risk assessment.

What should organizations using AI do?

Companies obviously do not have to analyze every dispute between a model provider and the US administration. But they should draw a few practical conclusions from this story.

First, they should know which AI tools their employees use. Without that, it is impossible to assess the impact of a sudden block, a model change or a change in terms of use.

Second, they should understand what data and processes they allow AI into. Using a model to improve the style of a text carries a different risk than using it to work with code, client documents, employee data or vulnerability analysis.

Third, they should have rules for using AI. Employees must know what may not be entered into tools, when a result requires verification and when the use of a model has to be consulted with a responsible person.

Fourth, companies should ask AI providers about security, data, availability, documentation, terms of use and the incident response process.

Fifth, they should have a contingency plan. If one AI tool suddenly disappears or changes its rules, the company should know which processes are affected and how to continue them.

Where does AI TrustCERT help with this?

AI TrustCERT helps organizations put this exact level of readiness in order.

The point is not for a company to know every technical detail of how Fable 5, Mythos 5 or other frontier AI models work. The point is for it to know how AI is used within its own organization.

AI TrustCERT helps move from scattered use of AI tools to a more controlled model of operation: with a register of tools, usage rules, employee training, acknowledgements, risk identification and a readiness report.

In the Fable 5 story, the most important thing is not whether a particular model returns tomorrow, today or in a few days. The most important question is whether a company would be ready for a similar situation of its own.

Would it know who uses a given tool? Would it know in which processes it is used? Would it know what data might have gone into it? Would it be able to decide quickly what to do after a change in availability?

If the answer is „no", then the problem is not about Anthropic.

It is about the organization.

Update: what will we watch after Fable 5 is made available again?

This article may be updated after access to Fable 5 is actually restored.

The most important things to watch are:

• how quickly the model returns to Claude.ai, the Claude Platform, Claude Code and Claude Cowork;
• when access is restored via AWS, Google Cloud and Microsoft Foundry;
• whether users will report more blocks of legitimate queries related to coding and cybersecurity;
• how the redirection of blocked queries to Opus 4.8 works;
• whether Anthropic will publish more details about the jailbreak assessment framework;
• whether access to Mythos 5 will be expanded beyond selected organizations in the US;
• whether the Fable 5 case will affect how other AI providers cooperate with governments when deploying the most advanced models.

This will be an important test not only for Anthropic, but for the entire frontier AI market.

Summary

Claude Fable 5 is coming back after a short but very loud block.

The model was suspended following a government directive related to export control and concerns about a safeguard bypass. Anthropic challenged the risk assessment, worked with the government and partners, strengthened its safety classifiers, and after the restrictions were lifted announced that it would make the model available again.

This story shows something more important than the return of a single model.

It shows that AI is entering a stage in which availability, security, regulation and geopolitics will affect the everyday use of tools in companies.

That is why organizations should stop looking at AI solely as a productivity application.

AI is becoming a part of the infrastructure of work.

And infrastructure has to be managed.