After a tense standoff with the US government, Anthropic’s most capable models are cleared for takeoff again — but only after the company agreed to bolt on an extra layer of restraint.
The Trump administration lifted export controls on Anthropic’s Claude Fable 5 after the company promised to extend an existing safeguard, according to two people familiar with the matter. The controls had effectively taken the model offline, so the reversal is a big deal for anyone who relies on it.
Here’s how the new guardrail works: if a user tries to access certain restricted capabilities, the request is blocked with a notification, and the query is quietly rerouted to the less-advanced Opus 4.8 model. That rerouting mechanism wasn’t invented from scratch — Anthropic already sent sensitive cybersecurity and biology requests to Opus 4.8. The update simply widens the net to cover a specific behavior flagged in a paper by Amazon.
What was that behavior? According to an analysis by Katie Moussouris, founder and CEO of Luta Security, users could sidestep a Fable 5 restriction with a clever bit of framing — asking the model to fix code rather than to identify security issues in it. Same result, different phrasing. Notably, most cybersecurity experts don’t consider this alarming. But once the administration caught wind of it, the confrontation with Anthropic escalated into full export controls.
Commerce secretary Howard Lutnick led the push to get the models back online, and his letter announcing the removal of restrictions on Fable 5 and Mythos 5 hinted at the deal. “Among other things, Anthropic has agreed to proactively detect and address security risks posed by the models,” Lutnick wrote. WIRED first obtained the letter.
The green light didn’t come purely on a handshake. Researchers at the Commerce Department’s Center for AI Standards and Innovation reviewed the safeguards and decided they were, in their words, sufficiently robust for now — a qualifier worth noting.
Still, Anthropic isn’t fully back in the administration’s good graces. Defense secretary Pete Hegseth has told advisers there’s no clear path to lift his February 28 order designating the company a supply chain risk, according to a person briefed on the matter. In other words, one door reopened while another stays firmly shut.
The episode is a revealing snapshot of how frontier AI is being governed in real time. A single prompt-engineering trick — one that specialists shrug at — was enough to knock a state-of-the-art model out of the market until its maker agreed to new controls. As models grow more capable, expect more of these behind-the-scenes negotiations over exactly which requests get answered, and by which version of the AI.
