US Lifts Export Curbs on Anthropic Fable and Mythos AI Models

Anthropic recently announced the US Commerce Department had removed export controls on its advanced Fable and Mythos artificial intelligence models. This decision follows a period of restricted access due to national security concerns, imposed less than three weeks prior.

Credit: CLAUDE.AI

Washington has intensified its oversight of new model releases, citing concerns that sophisticated AI could be exploited by military intelligence in China, Russia, or other countries of concern. An order issued earlier in the month compelled Anthropic to immediately limit access for Mythos 5 and Fable 5 to foreign nationals, leading the organisation to disable both models for all users. Nationality verification was not possible in real time.

A few days prior, the US government permitted Mythos 5's release, but only to approved US organisations. This model, designed to detect cybersecurity vulnerabilities, had previously been accessible to a wider array of companies in Anthropic's Glasswing programme. But Anthropic has now confirmed all export controls on these models are lifted, following new safeguard implementations.

Anthropic is now collaborating with the US government to broaden access for Mythos 5 to the wider network of Glasswing partners. Fable 5, with robust safeguards, is set to become available subsequently. The AI developer, which has experienced a challenging relationship with U.S. President Donald Trump's administration this year, also indicated a deepening collaboration with the US government, offering designated partners expanded early access to both models.

And the organisation is working with Big Tech firms, including Amazon, Microsoft, and Google, alongside other Glasswing partners. Their collective aim is to establish common standards for assessing and resolving AI "jailbreaks"-techniques designed to bypass security protocols. According to Reuters, US Commerce Secretary Howard Lutnick communicated Anthropic agreed to diligent cooperation on model protocols and malicious activity reporting. Mr Lutnick stated the department reserved the right to reevaluate conditions.

The earlier export control order stemmed from Amazon researchers reporting a method to circumvent Fable 5's safeguards, allowing it to identify software vulnerabilities and demonstrate exploitation. So Anthropic has since deployed a new safeguard, specifically blocking the behaviour outlined in the report. Any blocked requests will now be rerouted to its Opus 4.8 model.

While users might find this frustrating, Anthropic deemed it necessary for wider model availability. The company cautioned that complete immunity from jailbreaks for any AI model is "probably impossible." And it noted the potential for a "universal jailbreak" to emerge.

Increased scrutiny of AI models this month began with U.S. President Donald Trump's executive order, establishing a voluntary framework for developers to offer "covered frontier models" to the US government before public release. But rival OpenAI has also encountered restrictions. OpenAI announced a few days prior that it had postponed the full public launch of its GPT-5.6 model at the US government's request, limiting access to a small group of vetted partners.

Anthropic's relationship with the US government has been notably strained. The Pentagon earlier this year classified Anthropic as a "supply-chain risk," following its refusal to permit models for mass domestic surveillance or fully autonomous weapons systems. And both OpenAI and Anthropic have confidentially filed for US initial public offerings.

• The US Commerce Department removed export controls on Anthropic's Fable and Mythos AI models, imposed less than three weeks prior.

• The restrictions were due to national security concerns regarding potential misuse by foreign military intelligence.

• Anthropic implemented new safeguards and is collaborating with the US government and other Big Tech firms to develop common standards for addressing AI "jailbreaks."

• OpenAI has also faced similar government restrictions, leading to a delayed public launch of its GPT-5.6 model.

• Anthropic's ties with the US government remain complex, highlighted by an earlier Pentagon "supply-chain risk" designation.

Source: Reuters