For three days in June, the public could use a model called Fable 5 before the federal government pulled it.
Fable 5 wasn’t Anthropic’s most capable system; that was Mythos 5, a more powerful model the company has kept restricted. But Anthropic called Fable 5 “Mythos-class” and said the model outperformed anything it had previously put in front of general users, with particular strength in software engineering and scientific research. The company also said, without much hedging, that some of those same capabilities could be misused once the guardrails came off.
Three days after release, the government invoked national security and issued export controls barring access to any foreign national, including the Anthropic’s own foreign-national employees. Anthropic couldn’t certify compliance user by user, so it cut access for everyone.
Mythos and Fable won’t be the last models to unsettle a government. AI development has now reached a point where these systems are moving too fast for the regulators to keep up, and new capabilities appear before anyone has figured out how to handle the previous batch.
A paper out of
tries to chart where this goes. Titled “From AGI to ASI,” it traces a path from artificial general intelligence — long the field’s stated goal — to artificial superintelligence, the zone beyond where machines eclipse human capability. A year ago, that framing would have read as speculation; the brief release of Fable 5 gives it a concrete reference point.
Governments are built to deliberate, and deliberation takes time, while frontier AI runs on a different clock with new releases, benchmark jumps, and fresh agentic tooling arriving week after week. A regulator works in fiscal years, and AI labs now work in weeks.
Business Insider reports that the White House and Anthropic are now working out a framework to grade how severe a security flaw in a new model is, and to decide when a flaw warrants stepping in. Those rules didn’t exist when the Fable 5 situation called for them, and they’re being drafted only now.
The DeepMind paper lays out four ways the jump from AGI to superintelligence might happen, and they aren’t mutually exclusive: more of the same, only larger, with more compute and data poured into bigger models; a genuine algorithmic shift; recursive self-improvement, where AI starts doing the work of AI research; and large collectives of coordinating agents that add up to more than any single system. Any of these could already be underway, and more than one could be running at once.
The last possibility is the one that strains the current regulatory approach, in which the government waits for a technology to settle before studying its effects and writing rules around them. AI hasn’t settled, and it is increasingly an input into the next version of itself.
To their credit, the Google DeepMind authors avoid overselling. They don’t treat superintelligence as inevitable or imminent, and they catalog the obstacles: not enough data, ballooning resource costs, the chance that today’s neural-network approach hits a dead end, the plain difficulty of frontier research, the problem of getting a machine to form genuinely new concepts out of raw experience. Regulation and public backlash appear on the list too, as sources of friction. But friction slows the technology only if the institutions applying it can move at the technology’s speed, and the Fable 5 episode suggests they can’t yet.
The open questions remain: When does handing someone API access count as an export? A jailbreak that reads as an ordinary bug one week can look like a national-security event the next, and there’s no agreement on where the line sits. Who decides a model is too capable to ship, and on what evidence; the weights, the compute, the cyber and biology scores, who’s using it, where it runs?
This will keep happening, and the next confrontation may have nothing to do with cybersecurity. It could be a model capable of designing a pathogen. It could be one that quietly tunes a power grid or runs a persuasion campaign at scale. Each case will demand technical depth most agencies don’t keep on staff. It will also require legal authority that is murky at best, plus the kind of international coordination that tends to arrive late.
This could be the moment governments start building real evaluation capacity, such as pre-release testing. The likelier near-term path, though, is the reactive one, with agencies catching up to each model after it ships.
DeepMind’s own conclusion offers little reassurance. Preparing for what comes after AGI, the authors write, will take forecasting and benchmarking and continuous monitoring, plus the ability to turn that work into policy quickly, across labs and governments and the research community at the same time. They describe the task as navigating a “high-velocity technological trajectory,” a phrase that concedes how little anyone can do beyond watching closely and adjusting as the technology moves.
Fable 5 will fade from the headlines before long, but it’s a useful preview of the pattern. Over three days, a product launch became an export-control matter and a compliance scramble, and a geopolitical event. By the end, the company and the government were negotiating standards that should have been settled before launch. Cases like it will recur, and there’s little reason to expect the labs to slow down so the rule-writing can catch up.
