An AI export ban that backfires on defenders

On the evening of 12 June 2026, Anthropic started switching off the two most capable models it has ever shipped. We received the directive from the government today at 5:21pm (ET), the company said, and the order was sweeping: suspend all access to Fable 5 and Mythos 5 by any foreign national — inside or outside the United States, including Anthropic’s own foreign-national staff. The independent developer Simon Willison, who logged the cut-off to the second, summed up the reaction in three words: Well this is nuts:

The stated reason was national security. The government said it had become aware of a way to “jailbreak” Fable 5 — to get past the classifiers that block its most dangerous outputs. On the face of it, that sounds like exactly the kind of thing export controls exist for. Then you read what the jailbreak actually was.

The ‘jailbreak’ was just security work

Anthropic described the method as a narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws.

The jailbreak “essentially consists of asking the model to read a specific codebase and fix any software flaws.” — Anthropic

That sentence should stop anyone who works in security. Reading code and fixing its flaws is not an exotic attack technique — it is the daily grind of defence. As Snyk put it, it is automated code review and vulnerability remediation, the same job performed by static analysis, fuzzers, AI-assisted code review — a capability used every day by the defenders who keep systems safe. The same prompt that lets a model find a planted bug for an attacker lets a blue team find a real one before the attacker does.

This is the asymmetry at the heart of the row. Katie Moussouris — founder of Luta Security, and a veteran of the 2015 fight over the Wassenaar rules that tried, and failed, to export-control security tools — argues the order harms US cyber defence rather than protecting it. Determined attackers have other models, other tools, and time. The people actually cut off are the defenders who had wired a best-in-class reviewer into their workflow.

It is worse than a clean on/off switch, too. To comply without recalling the model entirely at first, Fable 5’s guardrails were tightened so aggressively that, IBM X-Force’s Valentina Palmiotti found, it now rejects any request that could be tangentially cyber related. A vulnerability analyst asking for help triaging a CVE gets refused; the capability is still in there, just walled off from the people with the most legitimate use for it.

Why this lands in the UK

For a British reader, the sharp end is the phrase any foreign national. Under the order, UK developers, agencies and in-house security teams are foreign nationals — so everyone here on Claude.ai, Claude Code or the API lost Fable 5 and Mythos 5 the moment it took effect, which Willison’s logs pin to 6:59pm Pacific. No notice, no transition window: a tool that teams had built into real workflows was simply gone within hours.

Anthropic, for its part, is complying while making its objection plain: we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.

What this signals

Two things are worth watching:

• The legal test. Anthropic’s challenge to the designation is the test case for whether “dual-use” AI rules can tell the difference between an attacker and a defender running the very same prompt. If they cannot, expect more defensive capability to be classified as a weapon and switched off by directive.
• The dependence risk. Snyk’s blunt lesson — that a single directive took a generally available product offline for its entire global user base within hours — is the clearest argument yet against building anything critical on one model, from one vendor, in one jurisdiction.

That second point is the one to act on now. If losing a single model would break a workflow you depend on, that is a risk to plan around — keep a second provider wired up and tested, not just bookmarked. The defenders who relied on Fable 5 found out the hard way that “generally available” can stop being available between one prompt and the next.