Some users keep Anthropic Mythos access despite US ban

TL;DR:

  • A handful of firms chosen to test Anthropic’s Mythos model have kept access despite a US shutdown order, Bloomberg reports.
  • Cybersecurity firms Cisco and Dragos confirmed retained access to Mythos Preview, while European agency ENISA was denied.
  • The episode adds nuance to a US export crackdown that has otherwise cut off foreign nationals.

The fallout from Washington’s clampdown on Anthropic’s most powerful model is proving uneven. Some organisations selected early to test Mythos have preserved their access to a preview of the system even after a US government order forced the shutdown of other versions, according to Bloomberg News. Anthropic had limited Mythos Preview to roughly 200 organisations — including the US government — under its Glasswing programme, after the model uncovered thousands of software vulnerabilities.

Selective access raises questions

A less powerful version of Mythos had been released publicly before being disabled under US export-control directives that suspended access for all foreign nationals on national-security grounds. Yet cybersecurity firms Cisco and Dragos confirmed to Bloomberg they retained access to the preview, and it was not immediately clear how Anthropic was deciding which Glasswing members kept it. Pointedly, the European cybersecurity agency ENISA — invited to join the programme before the US block — was told it would no longer be granted access.

That contrast matters for European and UK security bodies hoping to evaluate frontier models for defensive use. It also extends a saga that has rippled through British policy: the US order previously forced Anthropic to suspend Fable 5 and Mythos 5, prompting Sir Keir Starmer to seek a UK carve-out — a request Donald Trump rejected. The backdrop is Anthropic’s own rupture with the US government this year after it refused to let the military use its models for domestic surveillance and autonomous weapons.

Looking forward

For the UK, the lesson is uncomfortable: access to the most capable AI systems is now a lever of US foreign policy, applied selectively and at short notice. The fact that named American firms kept preview access while a European agency was cut off underscores how export controls can reshape who gets to probe — and defend against — frontier-model risks. It strengthens the case, already building in Westminster, for sovereign capability and trusted-partner arrangements that do not leave critical national-security tooling at the mercy of another government’s directives.

Share this article