US order to Anthropic to shut down key AI models misses too many critical points

The sudden dramatic US order to Anthropic to shut off access and disable AI platforms was given on the grounds of risks to national security. The order has been applied to Fable 5 and Mythos 5 in practice.

The stated administration theory is that there’s a way of jailbreaking Fable 5 to prevent it from identifying software vulnerabilities. Anthropic doesn’t agree at all and says the recall negatively affects hundreds of millions of users.  They add that their only advice from the administration about the matter was verbal.

Anthropic’s statement on the suspension spells out in torturous detail its position. They point out that they’re actually in favour of the government’s right to block unsafe developments but that they consider the order a “misunderstanding”.

Is this the right way to manage a real threat?

There’s a real tangle of truly thankless logic and priorities here. Fable 5 and Mythos 5 are very high-profile AI models with the ability to identify risks. Now, according to the Pentagon and others, are these two particular models suddenly big risks?

There are several aggravating issues involved for Anthropic in this suspension.

Fable 5 and Mythos 5 are both acknowledged leaders in AI models. They’re absolutely critical core business for Anthropic, and they are therefore directly related to the proposed IPO for Anthropic expected in the near future.

Anthropic has a lot of expensive skin in this game, and so do Anthropic’s many directly committed partners including major-leaguers like Google, Amazon, and Microsoft. Any sustained shutdown or failure to resolve these issues can only damage Anthropic’s bottom line and do collateral damage to these partners.

From a security and intelligence perspective, there are some oversize issues.

AI is unlike any other class of technology in security contexts. Development and exploitation of vulnerabilities can and do occur instantly. “Edicts from on high” are completely out of place here. It’s not like a bad actor would wait to act.

The realities of the new AI capabilities are far less ambiguous. Anthropic’s Claude Mythos is assessed by the AI Security Institute as an effective expert-level security exploitation asset. The cited effectiveness of Mythos as of April 2026 is stated as 73%, which is pretty convincing as a high-level threat. Mythos is effectively being positioned in public perception as an offensive cyber weapon.

Note the discontinuity between the order and timeframes for information about Mythos 5 and Fable 5. All of this information was available to the administration and the wider AI community at least two months before the suspension order.

This order could be very much an own goal for the administration. It’s easy enough to see the order as exceptionally naïve and surprisingly clumsy. Highlighting risks to yourself from your own technology isn’t good geopolitics. Why tell the world?

Still not addressing the real dangers

The administration is now doing something very like a forced version of the exact opposite of its original stated position on AI development.  The administration was vehemently and constantly opposed to AI regulation and had to be dragged screaming into some vestige of acceptance.

There are better ways of effectively managing AI security. Anthropic’s much-discussed oblique statements and actions for supporting AI regulation might well be the better and safer way of addressing these issues.

It there’s a real threat current to US security, Fable 5 and Mythos 5 are the only visible frontline first responders to that threat.

Scenarios include some pretty grim situations with far more questions than answers.

What if a cloned version of these AI capabilities hits vulnerable third parties which are now out of the loop for support from Fable 5 and Mythos 5?

Is it possible to even quantify the number of backdoor vulnerabilities for exploitation worldwide?  

Do these vulnerabilities allow indirect or direct access to US cybersecurity assets?

What are the safeguards and guardrails for Fable 5 and Mythos 5? Do they work, and has anyone tested them?

“Confidential computing” may be a stopgap, but can it do the job against frontier AI capabilities? It’s a very new approach only now being introduced and may have problems on older systems.

The world must be on the same page about AI security and stay on it

The world has gone far beyond the theoretical risks of AI and into the real threats. The need is for consistency and effective risk management. Rhetoric is meaningless. Pampering commercial interests at the expense of security is likely to be fatal.

Anthropic seems to have been trying to build in safeguards before deployment. Setting standards might be no more difficult than something like ISO standardization. Dull, but extremely useful with common parameters for global communications.

The current situation is a mess, actually amplifying risks without even being able to pretend to solve them. The world needs to do better than this.