Fable 5 Is Back, the Government in the Loop

The fable got a sequel, and a lot sooner than I'd have bet. Three weeks after pulling Fable 5 on Washington's orders, Anthropic has switched it back on. On the evening of June 30, the Commerce Department lifted the export controls that had hit Fable 5 and Mythos 5, and as of July 1 the consumer model is available worldwide again^[1]. I'd left the story on a question, how long a country can afford not to have a model of its own, and now the news answers first with a different one, trickier, about the terms of the return.

The story resumes, with strings attached

The comeback isn't a simple flip of the switch. Fable 5, the model open to everyone, returns worldwide, but Mythos 5, its guardrail-free version reserved for cybersecurity professionals, only reopens for a circle of approved US organizations, the government having signed off on June 26^[1:1]. Put plainly, the muzzled model goes global again while the unmuzzled one stays in American hands. The sovereignty I wrote about last month didn't wait long to make its point.

There's a subtler catch. To keep the incident from repeating, Anthropic redeployed Fable 5 with a fresh batch of classifiers meant to spot and block the cybersecurity requests it deems sensitive. The trouble is that the net is still coarse, coarse enough that chores as ordinary as writing or debugging code get bounced, for now, over to Opus 4.8, last generation's model^[1:2]. The company admits its own classifier snags plenty of perfectly legitimate requests along the way, and promises to sharpen it over the coming weeks so it can tell real misuse from a developer's ordinary curiosity. So Fable is back, just a notch more muzzled than before.

The accuser was Amazon

Last month I pieced the affair together from the little we had, a "rival company" that supposedly bypassed the model's protections. The veil has lifted, and, I'll admit, the name is worth a grin. The accuser was Amazon, whose researchers showed that feeding Fable 5 a codebase could get it to flag a string of software flaws and, in one case, to produce a rough bit of code to exploit them^[1:3]. It was Amazon's chief executive himself, Andy Jassy, who reportedly tipped off the White House^[2]. Washington's version is less flattering to Anthropic, mind you, since adviser David Sacks accuses the company of refusing to fix the flaw before the controls came down, while a Chinese group had reportedly already gotten hold of the model, against a backdrop of ongoing friction with Beijing^[3].

And Amazon is no ordinary rival. It's Anthropic's largest backer, the one that has poured billions into the company while raising its own house models on the side. Watching it trigger, with a quiet note to the authorities, the shutdown of the crown jewel it partly owns is quite the high-wire act. The punchline lands even better, since this same Amazon now sits among the partners drafting the framework meant to judge how serious a jailbreak is. The arsonist has picked up a fire hose.

Then comes the best part, the one that vindicates the suspicion I aired in June. Anthropic now says Amazon's demonstration was nothing special, since every model it tested produced the very same result, from little Haiku 4.5 up through the successive Opus releases, by way of OpenAI's GPT-5.4 and 5.5 or China's Kimi K2.7^[1:4]. The flaw wasn't in Fable, it was in the notion that a model able to read code could, by its very nature, spot flaws in it. They unplugged a tool for a knack every serious model on the market shares.

The state moves upstream

If the model came back, it wasn't for the price of a simple patch, but of a lasting rapprochement between the company and the government. Anthropic says it's expanding its collaboration with the US government on evaluating its models, which covers three concrete commitments: early access to models and their safeguards before public release, quick sharing of information on jailbreaks and misuse, and dedicated resources for joint research^[1:5].

What this arrangement sketches goes well beyond last month's incident. The state no longer steps in only after the fact, unplugging what worries it; it settles in upstream, in the engine room, where it's decided what a model will be allowed to know how to do. The "state secret" I described turns into co-management, quiet and permanent, between a private company and its government.

The move doesn't stop at Anthropic's door. With Amazon, Microsoft, and Google, the company is sketching a shared framework for gauging how serious a jailbreak is, along four criteria running from the real capability it hands an attacker to how easily that capability can be put to use^[1:6]. In parallel it's opening a HackerOne program so security researchers can flag the next flaws. The intent is fair enough, and it would be churlish to sneer at an industry finally trying to give itself some rules. Except that these rules are written by the same handful of American companies, watched over by the same government, for models the rest of the world will simply get to use.

Sonnet 5, the lesson learned

One last detail, and not a trivial one. The day before the announcement, on June 30, Anthropic shipped Claude Sonnet 5, a lighter model that aims at Opus 4.8's performance for a good deal less money^[4]. But the trait that catches the eye, given the moment, is that Sonnet 5 deliberately underperforms on cybersecurity tasks, its safeguards there switched on by default. June's lesson clearly stuck, and it turned into a product feature, a model built not to know how to do the very thing that might get it unplugged.

So Fable 5 is back, and I'd be lying if I said I wasn't glad, because getting a good tool back is always a small pleasure. Only the fable changed its moral along the way. It no longer says merely that a government can switch a model off with the stroke of a pen; it says a government can now pull up a chair at the table where the model is built, and stay there. The power is back on. The real question is who's holding the switch now.