Locking Down Autonomy
The effort to lock down access to autonomous AIs is less about security and far more about keeping the benefits of their emergence concentrated in the hands of incumbent elites.
In June 2026, the US government restricted access to frontier AI models. Here’s what happened;
Anthropic released two versions of the new model, each with significant improvements in operational autonomy.
Fable 5 (a general-use frontier model with guardrails) and
Mythos 5 (an unlocked version available only to Anthropic’s trusted partners).
Upon review, the government concluded that the autonomous cybersecurity capabilities of both models (Fable 5 via a jailbreak) posed a national security risk. For example, it could expose and fully exploit (up to 32 steps) cybersecurity vulnerabilities at a torrential pace.
The government immediately took action by imposing a “no foreign” (no outside-the-US use or use by foreign nationals) export restriction on the products.
Anthropic responded by shutting down all access to both products.
Since then, Mythos 5 has been released to a government-vetted list of 100 companies and organizations.
Fable 5 is still in negotiations with the government on ways to restrict access (all users will be required to submit IDs).
Why Now?
Three drivers made this shift, from open to restricted access, inevitable;
Renationalization.
Fear-mongering by the CEOs of big AI companies.
Rapid improvements in autonomy.
Let’s examine each in turn.
Renationalization
Globalization is in widespread retreat across the West, from immigration to economic competitiveness.
Globalized countries (hollow states, shells of laws/regul
ations/economic activity) are starting to act like nations again to avoid economic decline and social chaos.
Renationalization places the nation’s economic prosperity, social cohesion, and safety above the needs of corporations selling to the rest of the world.
With regards to AI development, the US government is using existing national security regulations in new, expansive ways to;
Prevent cutting-edge autonomous AIs from being used by rogue actors (states, groups, individuals) to wage cyberwarfare against the US, US companies, and US citizens.
Slow Chinese open-source model development (military and economic competition).
Maximize the benefits of frontier AI development for US citizens and companies.
Fear Mongering
The CEOs of many of the top AI companies have spent the last few years telling everyone that autonomous AIs are dangerous.
These claims are self-serving.
By hyping the danger of AI, these CEOs were prompting the government to take action to strictly limit and control the number of AI companies (aka pulling up the drawbridge on competition), particularly open-source models.
Furthermore, although most CEOs publicly claim they oppose national security access controls, they all want help shutting down state-sponsored model distillation attacks (to slow the pace of Chinese open-source development). Restrictions allow the NSA to help them with this.
Autonomy
Rapid improvements in AI autonomy made this move inevitable.
Model improvement is now laser-focused on enhancing AI autonomy (operating independently).
Frontier AIs can accomplish a full workday of tasks (~16 hours) with dozens of steps, all without oversight. Also, upon completing the task, the human user can make adjustments to improve it (a guided learning cycle in which the human provides orientation) and then launch it for the next task (OODA loop).
The best programmers can now comfortably oversee three to five autonomous AIs simultaneously (a shift from broader/deeper serial operations to parallel operations) and often claim an order-of-magnitude improvement (or more) in productivity (10-100x). We are going to see similar claims in every skill set/industry autonomous AI touches (from law to investing to marketing to construction to farming).
These superusers are quickly becoming national assets, not only for what they can do now, but also for what they will be able to produce in the future across everything from national security to economic competition. From a national perspective, if the gap between US superusers and the rest of the world can be locked in now, it will quickly widen as technology improves and these users vault down the learning curve.
Bait and Switch
These converging forces — the renationalization of frontier capabilities, the strategic use of safety arguments by incumbents, and the rapid improvement in operational autonomy — point to a reason for the crackdown that is far more important than export controls on frontier models.
