This Wednesday, June 10, 2026, the American company Anthropic published two major policy documents intended to frame the development of advanced artificial intelligence worldwide. Faced with progress described as exponential by researchers, the company believes that current political institutions, designed for a world that changes slowly, are no longer adapted. Its message is clear: it is urgent to act before model capabilities exceed our ability to control them.

Two frameworks to prepare for the age of powerful AI

The two documents released today have explicit names: the Advanced AI Framework and the Economic Policy Framework. Together, they form a coherent roadmap for governing a technology whose rise in power is now difficult to dispute.

The first framework addresses catastrophic risks linked to the most powerful models. The second focuses on how governments can prepare their economies and workers for the massive impact of AI on employment and wealth distribution.

“AI is progressing at exponential speed, and the political process was designed for a slower world.”

— Anthropic, Policy on the AI Exponential, June 2026

A blocking power for governments

The most radical proposal in the advanced framework is probably this: governments should have legal authority to block or deter deployment of models presenting a significant risk of catastrophic harm. This authority would go beyond what exists in current laws or pending legislative proposals in the United States.

In return, strict safeguards are planned to avoid abuses of power or the crushing of innovation. Penalties for violations would be civil, proportional to the company's annual global revenue, and would increase in cases of repeat offenses.

This framework would not apply to every company in the sector. It would target only developers of models trained with more than 10²⁵ floating-point operations (FLOPs), belonging to companies generating more than $500 million in AI-related revenue or spending more than $1 billion on R&D in the field.

The four catastrophic risks identified

Anthropic details four categories of risks in its document that justify strong regulatory intervention:

• Biological risk. The same capabilities that accelerate drug discovery can, without safeguards, be used to facilitate the creation of biological weapons or dangerous viruses. AI makes this threat both cheaper and more accessible.
• Cyber risk. Frontier models can now detect critical vulnerabilities at large scale in essential software. If these capabilities are used offensively, hospitals, power grids and critical infrastructure become potential targets.
• Loss of control risk. As AI systems improve, it may become much harder to control agents acting outside their developers' intentions — a scenario that has worried many researchers for years.
• Automated R&D. AI is now automating a growing share of AI research itself. This recursive acceleration mechanism could amplify the three previous risks in unpredictable ways.

Concrete obligations for frontier developers

Anthropic's framework proposes a set of requirements for companies developing frontier models:

• Mandatory transparency: publication of test results, system cards assessing capabilities and risks, and periodic reports on the overall risk posture.
• Independent evaluations: mandatory use of at least one qualified external evaluator, separate from the developer company.
• Enhanced security: protection of model weights and training infrastructure against malicious actors, including states.

Anthropic notes that several U.S. states — including California and New York — have already passed laws requiring transparency. But the company believes transparency alone is no longer enough: “The speed of acceleration means that transparency alone is no longer sufficient. Governments must play a more substantial role.”

A backdrop of record fundraising and imminent IPO

These policy proposals come amid spectacular growth for Anthropic. In late May 2026, the company announced a $65 billion Series H fundraising, bringing its valuation close to $965 billion. On June 1, it confirmed that it had confidentially filed an S-1 document with the Securities and Exchange Commission (SEC), the first step toward an imminent IPO.

In addition, the day before this policy publication, Anthropic launched its Claude Fable 5 and Claude Mythos 5 models, presented as the next generation of its systems for advanced coding tasks and intensive intellectual work. The Mythos Preview model had recently demonstrated its capabilities by discovering thousands of critical vulnerabilities in major operating systems and web browsers — concretely illustrating the cyber risks mentioned in the policy framework.

A different approach for federal and state governments

Anthropic is careful to specify that its framework is designed primarily for the U.S. federal government, while acknowledging that AI risks cannot wait for federal action. The company opposes any federal preemption of state laws unless the federal government adopts a law at least as robust as the framework proposed today.

This position illustrates an American political reality: without strong federal legislation, states play the role of first-line regulators for emerging risks — a role Anthropic wants to preserve while calling for national coordination.

And France in all this?

Although Anthropic's proposal is aimed first at the United States, its implications are global. The European Union, with its AI Act progressively applied since 2024-2025, has already laid similar foundations on transparency and risk categorization. But the debate on authorities' blocking powers — the ability to prohibit deployments before they cause harm — remains open in Europe as elsewhere.

France, which hosts major players such as Mistral AI, is closely watching these debates. The publication of this framework by one of the most influential companies in the global AI sector could accelerate European discussions on the strongest regulatory instruments.

The urgency according to Anthropic: act before it is too late

The tone of Anthropic's proposals is deliberately pressing. The company insists that model capabilities will continue to progress rapidly in the coming months, and that governance must keep pace. It urges policymakers to engage actively on these issues now — not in two years, not after a catastrophe.

This move is notable because it comes from a company that itself produces these powerful models. By proposing rules that would apply to its own products, Anthropic seeks to strengthen the credibility of its vision of responsibly developed AI — and to influence the international regulatory framework in a direction it sees as both safe and compatible with innovation.

The two full documents are available on Anthropic's official website. The debate is only beginning.