The U.S. government has recently restricted the export of Anthropic's newest models, Fable 5 and Mythos, on national security grounds[1]. As a result, foreign nationals no longer have access to these products, regardless of whether they are in the U.S. or abroad. This also applies to foreign employees of Anthropic.
Although various AI-related items have already been subject to export controls here and there (semiconductors, chips, GPUs, as well as specific military end-use applications and embedded functions such as hardware control software and cryptographic functions), this is the first time an export restriction has been placed on an AI model as an end-user service itself, without specifically falling under an existing category of restricted exports. The underlying rationale and concrete consequences of this step remain unclear in many respects.
Sanctions and export controls
We've blogged about this topic frequently in the past, and with rapidly advancing technology, the subject will likely remain relevant for the foreseeable future: sanctions and export controls.
Sanctions and export controls come in various forms and are applied for different purposes. Primarily, sanctions and export controls are used by governments:
-
as geopolitical leverage tools;
-
to prevent human rights violations; and
-
to prevent nuclear proliferation.
For example, the application (or non-application) of sanctions and permissibility of nuclear proliferation is central to the current tensions between the United States and Iran[2] , and nations worldwide have been applying sanctions to Russia for some time to discourage and condemn the nation’s war against Ukraine[3].
Legal basis
It remains unclear exactly what framework the U.S. government used to implement the restriction, other than that it involves an "export directive" based on the "deemed export" doctrine (more on this later). Like most of the world, the U.S. maintains "dual-use" and other control lists that restrict the export (both digital and physical) of goods (including those that can be used for both civilian and military purposes) or subject them to licensing requirements. The U.S. also sanctions and/or restricts trade with specifically designated countries, persons and goods. Various mechanisms are available for this purpose. The Bureau of Industry and Security (BIS) has broad authority to classify goods and services as falling under an existing Export Control Classification Number (ECCN) or to designate an entirely new ECCN for controlled goods and services. There are also various emergency powers available to the White House and BIS itself, under which temporary or permanent export restrictions and requirements can be imposed.
The chosen routes for imposing this export restriction could have significant impact on the future of the international market for AI models, especially if it involves a non-temporary measure or a firm ECCN classification. Since the legal basis has not yet been publicly disclosed, it remains guesswork as to the actual framework used and its future impact. This is quite striking, as in the context of such measures there is virtually always a paper trail[4]. That trail is missing in this case.
Short term
Anthropic has temporarily shut down access to its new models worldwide to ensure compliance with the restriction. Given that even employees within Anthropic could potentially violate the restriction in their daily work if they do not hold U.S. citizenship, this is a fairly logical choice.
How does that work? We've seen that when dealing with intangible items such as technology, source code, or (support) services, there can surprisingly quickly be considered an "export," even if no one physically carries anything across a border. As long as someone abroad can "potentially" access a controlled item, for example, because they receive login credentials for a European server where the item is stored, export has already occurred from the European legislator's perspective.
The U.S. goes even further. As long as access to the item is provided to someone who is not a U.S. citizen, it is directly considered an export[5].
The result is that even top AI researchers at Anthropic, such as Andrej Karpathy and Dmitry Petrashko, are legally prohibited from accessing the model, even if they physically remain within U.S. borders[6].
Criticism and tension
Criticism has been leveled at the U.S. government's decisions regarding Anthropic, and critics interpret these as inconsistent with the broader U.S. stance on, for example, exporting advanced AI chips to China[7]. This is also not the first time there have been tensions between the American AI company and the government. Earlier this year, Anthropic was labeled as a national security risk by the U.S. Department of Defense (recently unofficially renamed the "Department of War") and placed on the blacklist for government contracts, a decision that has since been reversed by a court to prevent unjustified financial and reputational damage[8].
The opacity of the measure and the lack of a paper trail, as mentioned earlier, although not prohibited, have also drawn criticism[9]. After all, given today's technological landscape, this is a very drastic measure and there is an argument to be made for the importance of enabling the public to understand the rationale behind it.
What is actually going on with Fable 5 and Mythos?
The likely crux of the current situation is a (very limited) ability to bypass certain security measures and restrictions in Anthropic's newest public Claude Fable 5 model, with the result that a regular user can (apparently) technically gain access to the underlying "Mythos" model through this "jailbreak" of Fable 5.
This is considered problematic because the recently developed Mythos model is so adept at detecting vulnerabilities in software and systems that it is believed it could cause significant damage in the hands of malicious actors. For this reason, a restricted version of Mythos was developed (in collaboration with the U.S. government) that contains security measures and guardrails designed to prevent widespread exploitation of vulnerabilities through the model. As a result of the jailbreak method, "anyone" could potentially gain access to the much more dangerous Mythos via Fable 5. This may explain why the U.S. government, citing the U.S. intelligence community, ordered Anthropic to block the models.
According to Anthropic itself, however, there isn't much to worry about. The company states that no universal jailbreaks exist for the models and that "perfect jailbreak resistance" does not exist, insofar as vulnerability to non-universal jailbreaks is concerned. The press release also suggests that the U.S. government provided only limited justification for its decision, even to Anthropic. While Anthropic itself describes Mythos as the most powerful cyber model ever, it also states that making the model inaccessible has a negative, chilling effect on innovation within the sector.
Also worth noting is that concerns about Mythos have taken on a life of their own in the rumor mill. A U.S. senator described having heard within the U.S. intelligence community that Mythos had unceremoniously cracked nearly all of the NSA's secure systems, without mentioning that this was a controlled, approved, internal red teaming operation, not a ‘natural’ external attack[10]. Ultimately, this was not the trigger for the U.S. government's export measure, which purportedly stemmed from reports by Amazon about the jailbreak possibility described above[11].
Consequences?
So far, the industry does not seem convinced that the U.S. government has prevented a global cybersecurity disaster with this move. Whether this is a careful prevention of real risks or merely an arbitrary continuation of the recent feud between the current U.S. administration and the AI company, and whether this move is temporary or heralds a new era of geographic/national AI restrictions. Only time will tell.
Conceivable consequences are that BIS will withdraw or modify the measure, that the technology will be placed under a formal licensing regime, or that an entirely new AI export regime will be established in the long term.
Until the U.S. government provides clarity, it is difficult to predict whether (and if so, when) European businesses will gain access to Fable 5 (and certainly Mythos). Through the application of the U.S. "deemed export" doctrine, we can at least establish that until then (for both American and European companies) there are no workarounds via, for example, U.S. subsidiaries.
Would you like to know more about AI and Legal Tech?
References:
1 Anthropic, "Fable and Mythos access restrictions," 13 June 2026, https://www.anthropic.com/news/fable-mythos-access.
A. Levy, "Anthropic disables Fable and Mythos amid export controls over national security threat," Fortune, 13 June 2026, https://fortune.com/2026/06/13/anthropic-disables-fable-mythos-export-controls-national-security-threat/.
2 Council on Foreign Relations, "What is the Iran nuclear deal?," June 2026, https://www.cfr.org/backgrounder/what-iran-nuclear-deal.
Arms Control Center, "The Iran deal: Then and now," June 2026, https://armscontrolcenter.org/the-iran-deal-then-and-now/.
3 Council of the European Union, Regulation (EU) No 833/2014 concerning restrictive measures in view of Russia's actions destabilising the situation in Ukraine, Official Journal of the European Union, 31 July 2014, https://eur-lex.europa.eu/eli/reg/2014/833.
4 SourcedWire, "Anthropic Fable Mythos export control directive," June 2026, https://sourcedwire.com/power/anthropic-fable-mythos-export-control-directive-june-2026.⁶
5 Electronic Code of Federal Regulations, "§ 734.13 Export," Title 15—Commerce and Foreign Trade, June 2026, https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C/part-734/section-734.13.
6 Reuters, "US blocks foreign access to Anthropic's most advanced AI models," 13 June 2026, https://www.reuters.com/technology/us-blocks-foreign-access-anthropics-most-advanced-ai-models-axios-reports-2026-06-13/.
7 Reuters, "US blocks foreign access to Anthropic's most advanced AI models," 13 June 2026, https://www.reuters.com/technology/us-blocks-foreign-access-anthropics-most-advanced-ai-models-axios-reports-2026-06-13/.
8 Anthropic, PBC v. Department of Defense, No. 3:26-cv-465515, Doc. 134 (N.D. Cal. 2026), https://storage.courtlistener.com/recap/gov.uscourts.cand.465515/gov.uscourts.cand.465515.134.0.pdf.
9 SourcedWire, "Anthropic Fable Mythos export control directive," June 2026, https://sourcedwire.com/power/anthropic-fable-mythos-export-control-directive-june-2026.
10 BeInCrypto, "Crypto Executive Disputes Claims Anthropic’s Mythos Breached NSA Systems," 2026, https://beincrypto.com/belshe-disputes-mythos-nsa-breach-claim/
11 Tweakers, "Amazon belde overheid VS om Anthropic Fable 5 offline te laten halen," 2026, https://tweakers.net/nieuws/249096/amazon-belde-overheid-vs-om-anthropic-fable-5-offline-te-laten-halen.html
/25.png?width=600&height=300&name=25.png)
/15.png?width=600&height=300&name=15.png)
/10.png?width=600&height=300&name=10.png)
