Anthropic Baked the Export-Restriction Saga Into Claude's System Prompt

Simon Willison published a curious quote from the Claude Opus 5 system prompt — the internal instruction that developers feed the model before it talks to users. The fragment isn't about the tone of answers or about formatting, but about a very specific episode from Anthropic's own recent history.

What happened to the models

According to the text of the prompt, the Claude Fable 5 and Claude Mythos 5 models were first released on June 9, 2026. As early as June 12, Anthropic suspended access to both models in order to comply with export restrictions imposed by the U.S. Department of Commerce. On June 30 the department lifted those restrictions, and on July 1 the company restored access. The details are laid out in a separate statement from Anthropic.

Why it's in the prompt

This whole sequence of events took place after the model had finished training. In other words, Claude can't know anything about the suspension from its own training data — its only source is precisely this internal note. Hence the need to spell out the facts right there in the instruction: without it, the model would either say nothing or start making things up.

What's notable is exactly how the guidance is worded. If a user asks about what happened, Claude is to confirm the facts calmly and matter-of-factly — and, in particular, not to deny that the suspension took place at all. The model is instructed to treat the export restrictions themselves like any other current political topic: to give a balanced and accurate account rather than share a personal opinion, and to refer people to the company's official statement for details.

It's separately noted that the situation may have changed since the note was written. So, if the model has the ability to search for information, it should check for up-to-date data, and if it doesn't, it should advise looking at Anthropic's website.

Why it's interesting

The episode is a good illustration of a practical problem with large language models: their knowledge freezes at the training cutoff date, and everything that happened afterward has to be topped up by hand. The system prompt turns into a kind of live patch to the model's memory — a way to neatly close a specific gap and, at the same time, to set the frame within which a delicate subject should be addressed in a restrained, fact-based manner. The headline under which Willison presents the quote sums up the goal succinctly: to keep Claude from giving wrong answers about the export-control situation.

Source: Simon Willison