Reflections on the Fable 5 Suspension and what it Reveals about AI Governance

I find the recall of Anthropic’s Fable 5 and Mythos 5 quite interesting. I am an AI user who cares deeply about safe and ethical AI, which made it important to find out why the project I was working on could no longer be built with the model producing the best outputs for it. I had only paused for the night, with plans to continue working the next morning, only to wake up to news that access to Fable 5 had been suspended.

According to Anthropic, the U.S government issued a directive ordering the suspension of access to Fable 5 and Mythos 5, for all foreign nationals within and outside the US, including any who were employed by Anthropic. To comply, Anthropic decided to disable the models for all customers, everywhere. The statement released by the company indicate that the government did not provide substantial evidence of a dangerous jailbreak. All they presented were concerns tied to a narrow jailbreak technique that occurred in other models built by Anthropic themselves, as well as other companies like OpenAI.

This raises questions about AI safety and governance that are worth exploring, which, in so doing, has led me to view the situation through three lenses:

• AI Improving Workflows
• AI being Safe
• Use and Misuse of Power

Let’s explore each.

AI Improving Workflows

AI models have become an active part of many workflows within and outside the tech ecosystem. Models like Fable 5 have become increasingly capable of handling complex reasoning, multistep planning, and code generation that change how individuals and businesses operate. These models are becoming core tools for researchers, entrepreneurs, product teams, scientists, policy analysts and people who are trying to build tools, move faster in their work, explore interests, and tackle hard problems.

This makes models like Fable 5 a key collaborator in designing architectures, refactoring large codebases, and identifying edge cases that would be quite costly to find through manual effort alone (just last week, in a bid to document better for the developers I work with, I used an AI model to surface edge cases to a flow I had worked on, enabling me factor in everything the developers would need). For many teams, especially in the Global South where I come from, access to such tools is less about convenience and more about levelling the playing field. It provides the opportunity to progress amidst constraints like limited staff, small budgets and funding options, and a generally weak infrastructure.

If we could have a behind-the-scenes peak at the aftermath of this sudden withdrawal, what we’d see is missed deadlines, canceled experiments, and teams scrambling to re-architect their workflows. Of course Fable 5 was only publicly available for 3-4 days, but people had begun to build with it and integrate it into their workflows. We can certainly fall back to the earlier models or even competitors, but these alternatives simply do not match the capability of Fable 5 in this instance.

AI models simplify, hasten, or improve workflows. They are integral to the future we are building. And if we can agree that AI is becoming a kind of digital infrastructure (think cloud computing, broadband, etc) then abrupt suspensions can be likened to shutting down sections of that infrastructure because of risks that have not been clearly articulated to the model provider, let alone the public. This doesn’t automatically make the action wrong; it just means actions like this require scrutiny, transparency, and where necessary, debate. No one power house should retain enough power to disrupt the lives of millions across the world.

AI Being Safe

On the other hand though, safety is not optional. As frontier models become more and more capable, the level of harm they can cause if misused or misaligned becomes more catastrophic. We’ve all seen the movies, we don’t want it replicated in real life. Anthropic in its press release on the suspension of Fable 5, was explicit about the limits of current safety techniques. They noted that “…perfect jailbreak resistance is not currently possible for any model provider” and that “every safeguard used in the industry is vulnerable to non-universal jailbreaks.”

As a result, Anthropic adopted what it calls “defense in depth” for Fable 5 and Mythos 5, with the goal of making jailbreaks either narrow in scope or very expensive to produce, combining these with monitoring systems that quickly detect and shut down successful attacks. From the outside, this looks like a company putting in the work to ensure AI is used safely.

From a government’s perspective, especially a government that receives early access to some AI models and runs national security evaluations, this evolving landscape is understandably concerning. If a jailbreak, even a narrow one, allows for cyber exploitation or any other catastrophic risk, the government may feel pressure to act first and explain later. This is understandably a rational stance, after all, I’ve never had to lead a state or country before.

But here’s an AI Governance question: how do we design safety frameworks that are robust and precautionary without overreaching, reducing access, and creating new inequities? How do we avoid building a world where “AI safety” is used as a weapon to further propagate existing forms of oppressions?

Use and Misuse of Power

This bring me to the third lens: power.

The U.S government did not, based on the statement by Anthropic, provide substantial evidence of any actual harmful incident caused by Fable 5 or Mythos 5. Instead, it went straight to ordering the restriction of access by foreign nationals. The choice of who to restrict may be the only reason I am writing this piece. It seems quite telling as the decision to restrict at all.

When a government can determine that certain groups of people - in this case, defined by nationality - should not access a particular class of knowledge or capability, it looks eerily similar to historical (and dare I say, ongoing) patterns of technological gatekeeping. I am reminded of Christine Mann Darden and the women of her time who were stuck punching numbers daily, while their male counterparts climbed up the ladder as engineers in 1960s NASA.

For builders in the Global South, access to frontier AI models is one of the few ways to compensate for gaps in our infrastructure, research funding, and even hardware. This is our opportunity to experiment with cutting-edge capabilities without having to build AI labs ourselves, because in all honesty, it would take a minute before we get to that point. Thus, when access is withdrawn disproportionately, especially for regions and groups that are already underrepresented in AI research and development, it raises difficult questions.

Is this an effort to curtail genuine security risks, or is it functioning, intentionally or not, as a mechanism that keeps frontier capabilities in the hands of a small set of actors? If Anthropic had turned off access for only foreign nationals, would we be forced into witnessing the early stages of a global AI divide?

What might AI Safety Become?

There is a lot of technical speak around AI safety: alignment, red-teaming, capability evaluations, jailbreaks, etc. These are crucial in the roles they play. But this episode with the Fable 5 recall is a reminder that AI safety is also political, economic, and social. It is about who gets to define what is classified as risk, who gets to access advanced capability, and whose interests are protected first when there is perceived threat.

In the book, Woman at the Edge of Time, Marge Piercy paints a picture of a future where there is no traditional centralised government, instead the society is organised through a highly decentralised model. Decisions are made through meetings, discussions and deliberation. Communities coordinate with one another when necessary and if one community wins a decision over the other, the winning community must host a feast for the other community. Authority generally flows from the local level upward rather than from a central government downward.

Going by this, it is safe to assume that if a member of the community, Mattapoiset, built an AI model that all other communities use and a perceived threat was found in this model, there is no one person or group of people who could decide to, and instantly cut off access. At worst, members of Mattapoiset would deliberate amongst themselves, arrive at a decision, and carefully communicate this with the other communities.

But there’s another future described in the book. In this one, centralised government remains. In fact, large corporations have effectively replaced or absorbed many functions of the government causing economic and political power to concentrate in the same hands. People’s behaviour, relationships, and opportunities are heavily monitored and managed and technology is used to maintain order and reinforce inequality.

We can say then, that one possible future for us is one where AI safety becomes a shared global project: governments, companies, civil society groups, and researchers co-develop standards, share evidence, and negotiate proportional responses to threats or jailbreaks. This future might still allow for export controls and usage restrictions, but they would be accompanied by clear justification, independent oversight, and mechanisms to ensure they do not reinforce existing inequalities.

Another possible future is that “AI safety” becomes a tool that powerful actors use to justify control with limited or no transparency, and minimal input from those most affected by the exertion of power. In this world, AI safety may lose its original intent: protecting humans from harm, and take on a different function: protecting specific geopolitical or economic advantages, even when the technical risks are ambiguous or modest.

The real future may lie between these two extremes and may even vary from one situation to the next. And this is precisely why we must be attentive to events like the suspension of Fable 5 and Mythos 5, and the precedent is sets.

Let us Engage More

What can we say we’re really looking at here? An applause worthy move towards safe AI, or an early sign that AI safety is being used to reinforce old inequalities and create new forms of exclusion?

At this moment, I am less interested in offering a definitive answer (would that Anthropic or I had one), and more interested in inviting more people to the conversation, especially among builders, policymakers, and communities outside of the traditional centres of AI power.

A few questions we can start asking publicly and repeatedly:

• What kind of evidence should be required before governments can order the restriction of access to widely deployed frontier models on national-security grounds?
• How can companies and regulators design procedures that minimise sudden, global disruptions to critical AI infrastructure while still responding responsibly to risks that emerge?
• How do we include voices from the Global South in decisions about what “safe enough” looks like, and about what stays and what gets suspended?

AI must be safe; this is clear. But how do we ensure that the banner of AI safety will be carried in ways that widen access, distribute opportunity, and reduce harm? How do we keep AI from moving us backward?

References

• D’Ignazio, Catherine, & Klein, Lauren F. (2020). Data Feminism. MIT Press.
• Piercy, Marge. (1976). Woman on the Edge of Time. Alfred A. Knopf.
• Anthropic. (2026). Statement on the US government directive to suspend access to Fable 5 and Mythos 5 **https://www.anthropic.com/news/fable-mythos-access**
• Anthropic. (2026). Claude Fable 5 and Claude Mythos 5. **https://www.anthropic.com/news/claude-fable-5-mythos-5**