In a striking sequence of events that underscores the breakneck pace and perilous stakes of frontier AI, Anthropic released its most capable model ever—then watched the US government move to restrict it within days. The episode raises urgent questions about whether the world’s most powerful AI systems are outpacing the institutions meant to govern them.
- Capability leap: Anthropic’s Fable 5 scored 53.3% on Humanity’s Last Exam, a significant jump over Claude Opus 4.8’s 45.7%, marking a new high-water mark for AI performance.
- Government intervention: US authorities restricted Fable just days after its public release, signaling a new willingness to intervene in real time.
- Dual-use danger: Anthropic previously withheld a similar model, Mythos Preview, because it was deemed too adept at finding cyber vulnerabilities for general release.
- Controlled access: A less-safeguarded variant, Mythos 5, was shared only with select trusted organizations, highlighting a two-tier access model.
A Record-Breaking Release—And a Rapid Reversal
On June 9, Anthropic unveiled Claude Fable 5 to the public, touting it as a major step forward in AI capability. The model’s 53.3% score on Humanity’s Last Exam—a notoriously difficult benchmark designed to test the outer limits of machine intelligence—eclipsed the 45.7% achieved by its predecessor, Claude Opus 4.8. For Anthropic, it was a milestone. For regulators, it was apparently a red flag.
Within days, the US government issued an order restricting the model. The precise scope of that restriction remains unclear, but the speed of the intervention is unprecedented. It suggests that federal agencies are now monitoring frontier AI releases in near real-time—and are prepared to act when they perceive a threat.
The Model Anthropic Deemed Too Dangerous to Share
The Fable 5 episode did not emerge in a vacuum. Months earlier, Anthropic had announced a model called Mythos Preview, which it deliberately withheld from general release. The reason: Mythos was judged too skilled at identifying cyber vulnerabilities—a capability that, in the wrong hands, could supercharge hacking, espionage, and infrastructure attacks.
Anthropic described Fable 5 as having capabilities similar to Mythos Preview. That admission alone should alarm security professionals. If Fable 5 is nearly as adept at finding software flaws as a model deemed too dangerous for public consumption, then its release—even with safeguards—represents a calculated risk. The company also made Mythos 5, a version of Fable stripped of strict bio and cyber safeguards, available to a small number of trusted organizations. That two-tier system raises its own questions: Who decides who is trusted? And what happens when a safeguard-free model leaks?
Anthropic’s Paradox: Pause Calls While Shipping Power
Adding to the tension, Anthropic has recently called for an “option to slow or temporarily pause frontier AI development.” The company argues that the industry needs mechanisms to hit the brakes if risks spiral out of control. Yet Anthropic itself continues to release ever more capable models at a rapid clip.
This contradiction is not unique to Anthropic—it reflects a broader industry dynamic where competitive pressure, investor expectations, and geopolitical rivalry make voluntary restraint extraordinarily difficult. The result is a classic arms race: each lab fears that if it slows down, a rival will surge ahead. The call for a pause, however sincere, rings hollow when the same company is pushing the frontier forward.
The Fable 5 saga exposes a systemic vulnerability at the heart of the AI boom: the gap between capability and control. Models are improving faster than regulators, security researchers, or even their own creators can fully understand. When a model can find cyber vulnerabilities better than most human experts, it becomes a dual-use weapon—equally valuable to defenders and attackers.
The most immediate risk is proliferation. Even with government restrictions, a model released to the public cannot be un-released. Weights, APIs, and outputs can be copied, reverse-engineered, or misused. The restriction may slow some uses, but it cannot eliminate them. Meanwhile, the existence of Mythos 5—a safeguard-free variant in the hands of a select few—creates a high-value target for nation-state hackers and insider threats.
There is also a workforce dimension. As AI systems grow more capable, they will increasingly automate tasks currently performed by cybersecurity analysts, researchers, and developers. That could lead to job displacement, but worse, it could lead to a dangerous over-reliance on AI for security decisions—systems that can be fooled, poisoned, or turned against their users.
Finally, the Fable 5 incident sets a precedent for government intervention in AI releases. That is a double-edged sword. On one hand, it signals that oversight is possible. On the other, it raises fears of politicized restrictions, regulatory overreach, or a patchwork of rules that stifle innovation while failing to address the deepest risks.
The Road Ahead: Governance at the Speed of AI
What happens next will define the trajectory of AI safety for years. Will the US government formalize a review process for frontier models? Will Anthropic and its peers accept binding limits, or will they relocate development to friendlier jurisdictions? Will the “trusted organizations” granted access to Mythos 5 be subject to audits, or will they operate in a black box?
One thing is clear: the era of releasing powerful AI models with minimal oversight is ending. The Fable 5 restriction is a warning shot—not just to Anthropic, but to the entire industry. The question is whether it will be heeded, or whether the next release will be even more capable, even more dangerous, and even harder to control.
The world is watching. And the clock is ticking.
Originally reported and sourced from Center for AI Safety.
