Anthropic Proposes AI Regulatory Framework: Government Should Have Authority to Block High-Risk Models
Claire Weston
Anthropic on Wednesday published a regulatory framework arguing the US government should be legally empowered to block catastrophically risky AI models — paired with a civil-penalty regime tied to global revenue. The proposal goes beyond existing law and any bill currently before Congress.
What exactly does Anthropic want the government to do?
Anthropic calls for legislation granting the government a new power: when a third-party assessment finds a model poses unacceptable risk, the government can block or restrict its deployment.
This means → not voluntary disclosure, but mandatory pre-deployment clearance — fail the test and the model cannot ship.
The company also proposes a fine structure tied to global annual revenue, with penalties escalating for repeat violations.
Who gets regulated? Where is the threshold?
The framework applies only to models trained with more than 10²⁵ FLOPs — a measure of the computing power needed to train a model — where the developer's AI-related revenue exceeds $500 million or AI R&D spending tops $1 billion.
In plain terms = this line catches only the largest models from a handful of frontier labs; the vast majority of AI companies and open-source projects fall well below it.
Assessments cover four specific risk domains: bioweapons, cybersecurity, AI-system autonomy, and automated R&D — conducted by independent bodies or authorized private organizations.
Why is Anthropic using its own model as evidence?
Anthropic previously stripped cybersecurity-related capabilities from the public release of its Mythos model, granting full access only to select partners.
CEO Dario Amodei cited Mythos as his central exhibit, calling it a "perfect illustration of AI's extraordinary capabilities and potential dangers."
This reflects a deliberate strategy: run strict internal processes first, then frame those practices as the reference standard for the industry — and push legislation that makes them everyone's obligation.
How far is this from where the White House stands?
The AI executive order Trump signed earlier this month asks only for voluntary developer access — no mandatory approval gate.
Amodei has already criticized the order publicly as too weak.
This means → a substantial gap separates Anthropic's framework from current White House policy — bridging it from "voluntary cooperation" to "government veto power" requires an act of Congress.
Genuine safety push — or building a regulatory moat?
Anthropic and rival OpenAI have both announced plans to pursue an IPO. Releasing a dense regulatory blueprint on the eve of listing invites scrutiny of timing alone.
The framework's three obligations — safety testing, public disclosure of results, independent assessment — closely mirror processes Anthropic already runs internally.
Put simply = if this standard becomes law, Anthropic is nearly pre-compliant, while the compliance burden lands on competitors. Whether federal legislation gains real traction in Congress is the test of whether this framework is a genuine public proposal or a competitive instrument.
Content is for reference only, not financial advice.