When an artificial intelligence maker suspends access to its most advanced model for nearly three weeks, it’s worth understanding what happened and what lessons it leaves behind. In late June 2026, Anthropic announced that it was restoring access to its Claude Fable 5 and Mythos 5 models after the United States government lifted export controls applied on June 12. Behind that news lies an issue that directly affects any private company that uses or wants to use AI: AI security and how to choose a provider wisely, without being at the mercy of every headline.
At Tisa we don’t sell a brand out of dogma. We are a Microsoft partner and we work with AI applied to business, but our job is to help you decide. That’s why we’ll tell you what happened, in plain language, and what you should take away for your company.
What happened with Claude Fable 5, according to Anthropic
According to Anthropic’s statement, on June 9 it launched two models that share the same foundation: Fable 5, with the strictest safeguards the company says it has ever applied, and Mythos 5, with fewer restrictions and available only to a small group of trusted partners (the Glasswing program) dedicated to defensive cybersecurity.
A few days later, Amazon researchers documented a way to bypass Fable 5’s protections: they got the model to identify software vulnerabilities and, in one case, to generate code showing how to exploit one of them. In response to that report, the U.S. government applied immediate export controls. Since there was no reliable way to verify each user’s nationality in real time, Anthropic chose to suspend access to both models worldwide.
Anthropic itself qualifies the scope: its tests confirmed that other less capable models—including earlier versions of Claude, GPT-5.5, or Kimi K2.7—could identify the same vulnerabilities. In other words, this was not a unique offensive capability. Even so, the company trained a new security classifier that, it claims, blocks the described technique in more than 99% of cases, at the cost of also rejecting more legitimate programming requests.
Classifiers and "jailbreaks," in plain terms
Two terms come up throughout this story and are worth understanding, because you’ll see them more and more.
A security classifier is a small automatic system that monitors requests to a model and blocks those that seem dangerous. Anthropic explains that it works with a deliberate "safety margin": it prefers to block some probably harmless requests rather than let a dangerous one through. The user experiences it as a model that sometimes refuses to respond to something reasonable.
A jailbreak is a way to trick those protections with clever instructions so that the model does something that should be blocked. Anthropic acknowledges that no model is completely immune and proposes, together with Amazon, Microsoft, and Google, a common framework for measuring the severity of each jailbreak according to four criteria: how much extra capability it gives the attacker, how many tasks it works for, how much effort it takes to turn it into a real attack, and how easy it is to find.
The underlying idea is sensible: if the whole industry measures severity in the same way, it’s easier to prioritize, communicate risk, and react quickly.
AI security: what lessons it leaves for your company
It might seem that this only concerns large labs, but AI security is already a business variable for any SME. Three concrete lessons:
1. A model’s availability is not guaranteed
A regulatory change cut off access to a cutting-edge model for nearly three weeks. If your critical process depends on a single model from a single provider, an outage like this stops you. The practical recommendation is to design your solutions to be able to switch models with minimal disruption, and not to tie essential processes to a single piece.
2. Safeguards have an operational cost
The statement itself admits that the new classifier rejects more legitimate code and debugging requests. Translated to your day-to-day: more protection usually means more "false positives." When evaluating an AI tool, ask not only what it can do, but how often it blocks valid work, because that affects your team’s productivity.
3. Without data governance, AI doesn’t perform
This whole discussion revolves around what information goes into and out of a model. For a company, the first line of defense is not someone else’s classifier, but having clarity about its own data: what is sensitive, who accesses it, and for what purpose. That is data governance, and it’s the foundation on which any AI use case rests.
How to bring AI down to earth wisely (and not because it’s trendy)
Our experience with private companies and SMEs—retail, distribution, construction, food industry, hospitality, or sports centers—tells us that AI performs when it’s applied to a concrete and measurable use case, not for the sake of "having AI." Before chasing the latest model in the headlines, it makes more sense to answer business questions: which process costs you hours every week? What data do you have organized to feed a solution? What savings or improvement do you expect?
At Tisa we accompany that journey step by step: advising to identify the relevant data and create governance protocols, data analysis to decide based on information rather than intuition, and the development of a use case tailored to your business. And we do it on a solid technological foundation: Microsoft Dynamics 365 Business Central as the ERP and Power Platform to extend and automate, with AI added on top when it brings real value.
The advantage of such an approach is twofold. On the one hand, you choose technology based on business criteria—total cost, sector fit, integration, and support—not on dogma. On the other, if tomorrow a provider changes the rules or suspends a model, your company is not left paralyzed.
Conclusion
The Claude Fable 5 episode confirms that the AI market moves fast and that regulation is beginning to set the pace. For your company, the lesson is neither to panic nor to ignore it: it’s to choose sensibly, organize your data, and support each project with a use case that has a clear return.
Want to explore where AI fits in your business without hype and without strings attached? At Tisa we’ve been helping companies modernize since 1987. Write to us at info@grupotisa.com, call us at (+34) 971 305 885, or visit grupotisa.com and we’ll give you a no-obligation assessment.