Anthropic’s most powerful AI models are back online, but the 18-day freeze that preceded their return exposed just how quickly government intervention can shut down frontier technology when security concerns surface.
The US government removed export controls on Claude Fable 5 and Claude Mythos 5 on June 30, clearing the way for Anthropic to restore access starting July 1. Fable 5 returns globally across all Anthropic platforms, while Mythos 5, a variant with fewer built-in safety restrictions, is being rolled out only to US organizations that received government approval on June 26. The distinction matters: same underlying model, different guardrails, and now different access tiers.
The trigger for this whole episode dates back to June 12, when Amazon researchers reported they had found a way to bypass Fable 5’s safeguards. The technique, known in security circles as a jailbreak, coaxed the model into identifying software vulnerabilities and, in at least one documented case, producing code that demonstrated how an exploit could work. That finding prompted the government to slap immediate export controls on both models.
Export controls are rules limiting which foreign nationals can access a given technology. The problem for Anthropic was practical: the order took effect immediately, and the company had no real-time mechanism to verify every user’s nationality. Rather than risk violating federal rules, Anthropic pulled the plug for everyone.
A Security Finding That Rippled Into Crypto Markets
The suspension did not stay confined to AI circles. Traders had been wagering on Anthropic’s value through a pre-IPO perpetual contract listed on Hyperliquid, the onchain derivatives exchange that has carved out a niche in speculative instruments tied to private companies. When Anthropic went dark on June 12, that perpetual dropped roughly 3.7%.
The logic was straightforward. Anthropic has been widely expected to pursue a public listing at some point, and any regulatory entanglement that freezes access to flagship products prompts skepticism about timing. Investors betting on an IPO timeline suddenly faced uncertainty about whether export-control drama would become a recurring pattern.
Hyperliquid’s pre-IPO perpetuals are futures contracts with no expiry date, letting traders take long or short positions on a company’s implied valuation until it actually lists. The 3.7% move might seem modest compared to the swings common in crypto tokens, but for a contract tracking a private company valued in the tens of billions, the signal was clear: regulatory risk just became more real.
Our prior coverage explored how Claude Fable 5’s powerful coding abilities raised alarms among security researchers even before the export freeze. The Amazon jailbreak discovery validated those concerns. When an AI model can not only identify a vulnerability but also produce working exploit code, the DeFi sector in particular has reason to pay attention. Smart-contract audits are only as good as the auditors, and a sufficiently capable model could, in theory, compress the timeline from vulnerability discovery to exploit deployment from weeks to hours.
That is exactly the scenario security professionals warned about. The fact that it took a jailbreak (a deliberate circumvention of safety measures) to get there is cold comfort. Jailbreaks circulate. They get refined. And once a technique proves viable, blocking it becomes a game of whack-a-mole.
How Anthropic Is Changing Its Relationship With Washington
The export freeze lasted 18 days, from June 12 through June 30. What comes next may matter more than the freeze itself.
Anthropic announced it is deepening its collaboration with the US government in several ways. First, designated federal agencies will now receive early access to frontier models and their built-in safeguards before public release. Think of it as a security audit by regulators before the product ships.
Second, Anthropic is sharing threat intelligence with government counterparts. When the company discovers a jailbreak or a misuse pattern, that information flows to agencies tracking national-security risks.
Third, Anthropic is working with Amazon, Microsoft, and Google on a common framework for scoring how dangerous a given jailbreak is. The idea is to create an industry-wide standard rather than leaving each company to assess risks in isolation. If one firm rates a jailbreak as critical while another rates the same technique as low-risk, regulators have no consistent basis for action. A shared scoring system, assuming the companies can agree on criteria, would change that.
These arrangements tie into an executive order on AI security signed on June 2. The order itself signals a policy shift: frontier AI releases are now expected to undergo some level of government review, at least for models that hit certain capability thresholds. That is a departure from the relatively hands-off approach that prevailed during the earlier boom in large language models.
Supporters will call this responsible practice. Critics will argue that giving Washington a gatekeeper role over which models ship introduces bottlenecks and potential political interference. Both perspectives have merit, and the tension is unlikely to resolve anytime soon.
The Crypto Angle Nobody Asked For (But Now Has to Consider)
Anthropy is privately held, so its stock is not directly tradable on public exchanges. But the existence of derivatives like Hyperliquid’s pre-IPO perpetuals means the company’s fortunes are already reflected in crypto-adjacent markets. That creates a feedback loop.
When Anthropic runs into regulatory trouble, traders betting on its implied valuation take losses (or gains, if they were short). Those moves, in turn, influence sentiment around AI-linked crypto narratives. The AI sector in crypto, tracked on our sectors dashboard, includes tokens tied to decentralized compute, AI model marketplaces, and inference networks. When frontier AI companies face constraints, the decentralized alternatives start to look more appealing as hedges against regulatory capture.
The Fable 5 jailbreak also has direct implications for DeFi security. If a model can generate exploit code, then anyone with access to that model (or a jailbroken version circulating in less-regulated corners of the internet) can accelerate their attack research. Smart-contract auditors have already begun incorporating AI tools into their workflows. The question now is whether malicious actors can use the same tools faster.
Consider the math. A typical smart-contract audit takes weeks and costs five to six figures. An AI model that can read Solidity, identify reentrancy bugs, and suggest exploit vectors could compress that process dramatically. The firms with the best models become either the most effective defenders or, if those models leak, the most effective attack surface.
This is not hypothetical. Earlier this year, Anthropic’s Opus 4.8 model uncovered an infinite mint bug in Zcash that had gone undetected for four years. That discovery crashed ZEC by 38%. The vulnerability was patched before exploitation, but the episode showed what happens when AI-level pattern recognition meets blockchain codebases that have been live for years without exhaustive audits.
What the Mythos 5 Restrictions Actually Mean
Fable 5 is back for everyone. Mythos 5 is not.
Mythos 5 shares the same underlying model architecture as Fable 5, but Anthropic built it with fewer safety restrictions. The use case was likely research and internal testing: a version of the model that does not refuse as many requests, allowing developers to probe its limits and understand its failure modes.
That is exactly why the government treated it differently. A less-restricted model is, by definition, easier to jailbreak or misuse out of the box. The June 26 approval for certain US organizations means Mythos 5 is now available to entities that passed some kind of vetting process. Anthropic did not disclose the criteria.
For the broader market, the distinction matters. Fable 5’s global return means API access resumes for developers everywhere. Mythos 5’s restricted rollout means the most capable (and least guarded) version of the model remains behind a government-approved gate. Whether that gate holds is another question.
Looking Ahead: A Tighter Loop Between AI Labs and Regulators
The Anthropic episode is a case study in how fast things can move when national-security concerns intersect with commercial AI. Amazon found the jailbreak, the government applied controls within days, and Anthropic had to make a binary choice: suspend access for everyone or risk federal liability. It chose the former.
The 18-day outage cost Anthropic some trader confidence (at least 3.7% worth on Hyperliquid), raised fresh questions about IPO timing, and forced the company into a closer relationship with federal agencies. Whether that relationship proves advantageous or constraining depends on perspective.
For crypto participants, the takeaway is practical. AI models capable of generating exploit code are now subject to the same export-control apparatus that governs advanced semiconductors and defense technology. That does not make the models less capable; it makes access to them more fragmented. US organizations get one level of access, the rest of the world gets another, and the underground economy of jailbroken models continues regardless.
The derivatives market on Hyperliquid provided real-time pricing of that regulatory risk. When controls dropped, the perpetual dropped. When controls lifted, the contract presumably recovered (though Anthropic has not yet disclosed updated trading data). The mechanism is blunt but honest: traders priced in the uncertainty, and the price moved.
Meanwhile, Anthropic’s new collaboration framework with Amazon, Microsoft, and Google will bear watching. A shared standard for scoring jailbreak severity could become a regulatory baseline: if the industry agrees that a given technique rates as high-risk, governments may feel more justified in restricting access. If the standard is too loose, it becomes a fig leaf.
The June 2 executive order on AI security points toward a future where frontier models require some form of government sign-off before release. That is a significant policy shift, and the Fable 5 freeze is an early illustration of what it looks like in practice. Models do not simply ship anymore. They ship after regulators have had a look.
For the crypto sector, which has spent years arguing for permissionless innovation, the contrast is instructive. AI labs are now operating under constraints that would be unthinkable in decentralized protocol development. Whether that constraint produces safer technology or simply pushes risk into less-visible corners remains to be seen.
Fable 5 is back. Mythos 5 is back for some. The government is now part of the development process, not just the enforcement apparatus. That is the new landscape, and anyone building at the intersection of AI and crypto needs to account for it.
Related Reading
References
Disclaimer: This is journalism, not investment guidance. Crypto is risky. Make your own informed decisions.
