Three open fronts after the clash between Anthropic and the U.S. government over AI safety

The article by James O'Donnell, published in MIT Technology Review on June 22, 2026, analyzes the most recent clash between Anthropic and the U.S. government, an episode that has opened deep debates about national security, technological sovereignty, and competition with China in the field of artificial intelligence. The author structures his analysis around three major axes that, in his view, deserve close monitoring in the coming weeks and months.

**The trigger: Mythos, Fable, and government intervention**

It all began in April 2026, when Anthropic publicly announced that it had developed an AI model called Mythos, so advanced in generating and understanding code that, according to the company itself, it could pose a global threat to cybersecurity. To demonstrate the real scope of the risk, Anthropic granted restricted access to a small group of cybersecurity experts, allowing them to evaluate the model's capabilities under controlled conditions.

Subsequently, the company released to the public a modified and supposedly safer version, dubbed Fable, on Tuesday, June 9, 2026. Just days later, that same Friday, the U.S. federal government declared that Fable posed a threat to national security and imposed export controls on it. Faced with this pressure, Anthropic revoked access to both models—Mythos and Fable—a few hours after receiving the government notification.

**Amazon's ambiguous role**

One of the most striking details of the episode, according to O'Donnell, is that it was Andy Jassy, Amazon's CEO, who alerted government officials that Fable would be dangerous. The detail is not minor: Amazon is simultaneously one of the main investors in Anthropic and a company that develops its own competing AI models. This dual status creates an obvious tension of interests that the article points out without drawing definitive conclusions, but which invites reflection on who defines the limits of security and with what motivations.

**The irony of the 'doomer' move**

The author observes with some irony that the so-called 'doomers'—activists and researchers who have spent years warning about the catastrophic risks of AI and demanding firm state intervention—have finally gotten what they were asking for: a drastic government action. However, that intervention did not come over a bioweapon or a runaway autonomous AI, but in response to a model that, in essence, is very good at programming. And the result, O'Donnell argues, looks more like a superficial reaction than an articulated, coherent security plan.

This point is especially relevant to the agentic AI ecosystem: if governments begin applying export controls to AI models for their coding capabilities, the impact on the development of autonomous agents—which depend heavily on generating and executing code—could be considerable.

**First front: the erosion of trust in American AI companies**

The first major consequence the article points to is the impact on the international perception of American technology companies. The possibility that the U.S. government could, from one day to the next, impose export controls that nullify access to an AI model sends an alarm signal to any company or organization in the world that depends on models hosted on American servers or developed by American companies.

French politician Bruno Retailleau called the episode a 'wake-up call' and presented it as an additional reason for Europe to accelerate the development of its own AI infrastructure. Many other European leaders joined this discourse, evoking the possibility of turning Paris into a new Silicon Valley. However, O'Donnell warns that this vision collides with an uncomfortable reality: China.

**The Chinese problem: open, cheap models with no gatekeepers**

Open-source models from China are, according to the article, highly capable, extraordinarily cheap, and can be downloaded to run on any organization's servers without needing to depend on any company or any government. This feature makes them enormously attractive to companies—both in the United States and Europe—that do not want to risk a U.S. government decree cutting off access to their AI tools from one day to the next.

But that same absence of gatekeepers and rules makes them equally attractive to malicious actors: precisely the kind of cybercriminals Anthropic hoped to keep at bay by building security measures into its models. The paradox is notable: by blocking access to a model with built-in security safeguards, the government could be pushing users and companies toward alternatives with no controls whatsoever.

The article mentions the case of the Chinese startup Zhipu, whose shares have soared, as an indicator of where the market could move if this trend continues.

**The question no one wants to ask: will the U.S. ban the use of Chinese models?**

O'Donnell poses a question that, in the current context, is perfectly logical though politically delicate: if American and European companies begin migrating en masse to Chinese models to avoid Washington's restrictions, could the U.S. government take the next step and declare that using Chinese-origin AI models poses a threat to national security? The author does not rule out this possibility, and given the current regulatory climate, neither should anyone who closely follows the intersection of politics and technology.

**Second front: the risk of leaving the country more vulnerable, not less**

The second major axis of analysis addresses a security paradox that prominent cybersecurity experts have already publicly highlighted through an open letter to the government: blocking access to Anthropic's models could make the United States more vulnerable to cyberattacks, not less.

The reasoning is as follows: cybersecurity researchers were using Anthropic's models precisely to prepare defenses and anticipate possible attack vectors. By cutting off that access, defenders are deprived of a valuable tool. Moreover, the letter's signatories argue that Anthropic's models are not significantly more dangerous than other leading models already widely available on the market.

This connects directly with the debate over nonproliferation applied to software. The concept of nonproliferation was developed to control physical materials such as the uranium used in nuclear weapons: scarce, difficult-to-produce elements that can be controlled by tracking their supply chain. Applying that same logic to software models—which can be copied, modified, and redistributed at near-zero cost—is, to say the least, problematic from a technical and legal standpoint.

O'Donnell also notes that it is far from clear that granting access to Fable counts as an 'export' in the legal sense of the term, suggesting that the government ban might not survive serious judicial scrutiny. This legal uncertainty adds another layer of complexity to the episode.

**Third front: the role of Congress and regulatory pressure**

The third element the article invites us to watch is the reaction of the U.S. legislative branch. The most recent precedent is illustrative: after the previous clash between Anthropic and the government over how the Pentagon could or could not use its models, a series of new bills were introduced aimed at defining the limits of military use of AI.

Currently, the actors with the most influence over how AI is deployed and regulated in the United States are the tech companies themselves and the White House. Congress has long been debating regulations on the use of chatbots by minors and is far from reaching consensus on the scope that government oversight of AI model security should have. However, each drastic action by the executive increases the political pressure on lawmakers to take action.

**The Trump administration's inconsistency as a background variable**

The article concludes by noting the difficulty of making predictions in an environment where the administration's stance toward AI changes frequently and without apparent logic. When President Trump came to power, he eliminated the restrictive regulatory framework that existed for the safe development of AI and promised to let tech companies operate freely. Yet the same White House has labeled Anthropic—the most valuable AI startup of the moment—a national security risk on two separate occasions: once in spring and again in summer of 2026.

The question O'Donnell leaves open is obvious: what will autumn bring? The lack of consistency in the U.S. government's AI policy is in itself a systemic risk for the ecosystem, because it generates uncertainty both for the companies that develop models and for those that integrate them into their products and services.

**Implications for the agentic AI ecosystem**

From the perspective of Manuel's agentic AI newsletter, this episode has direct implications worth highlighting. Autonomous agent systems—which carry out complex tasks by chaining actions, tool calls, and code generation—depend critically on high-capacity foundation models. If those models can be blocked from one day to the next by government decision, the reliability of any agentic system built on them is called into question.

The tension between the need for powerful models (which are the ones that work best as reasoning engines for agents) and the risk that those same models become subject to regulatory restrictions is one of the structural dilemmas the sector will have to resolve in the coming months. The possible migration toward Chinese open-source models as an alternative poses its own problems: the absence of guaranteed enterprise support, uncertainty about the quality of security safeguards, and potential geopolitical implications for the organizations that adopt them.

**Conclusion**

The MIT Technology Review article offers no definitive answers—nor does it aim to—but it does clearly articulate the three major fronts that will define the evolution of this conflict: the geopolitical reorientation of companies and governments toward non-American alternatives, the security paradox generated by blocking defensive tools, and the possible acceleration of legislative regulation in a Congress that until now has lagged far behind events. It is required reading for anyone who closely follows the intersection of politics, security, and advanced AI development.