Anthropic CEO Dario Amodei published a policy article titled 《Policy on the AI Exponential》 on June 11, for the first time publicly calling on the government to intervene as the company’s top leadership, and setting regulations for cutting-edge AI models. Amodei said that the Claude Mythos Preview can already identify high-severity vulnerabilities, as the basis for legislation.
Three Major Policy Claims: Confirmed Quantitative Standards and Legal Framework
Claim One: Mandatory Third-Party Testing
Quantitative Thresholds (any one triggers): training compute exceeding 10^25 FLOPs, or annual AI revenue exceeding $500 million, or AI R&D spending exceeding $1 billion. For companies whose models meet the thresholds, the models under their control must pass independent audits before they are listed; the tests focus on four risk areas: cybersecurity, biological weapons, runaway AI systems, and automated R&D that can accelerate the preceding risks. In the original text, Amodei confirmed that under this framework, the government has legal authorization to block, delay, or discourage model deployment.
Claim Two: AI as a Critical Cybersecurity Infrastructure
Require leading developers to protect “model weights” from theft by external attackers or internal personnel, and establish a statutory reporting channel for “model distillation attacks.” Amodei cited that Claude Mythos Preview can actively identify major high-severity vulnerabilities in major operating systems as the technical basis for this claim.
Claim Three: Openly Acknowledge Structural Labor Replacement
The framework states explicitly that if AI reaches the predicted level of capability, it will be a “complete substitute for the workforce” rather than an assistive tool for productivity. The framework models scenarios including unemployment rates of 5%, 10%, and more extreme cases, and advocates corresponding mechanisms such as wage insurance, a Universal Basic Income (UBI), and a sovereign wealth model.
Anthropic’s Confirmed $350 Million Policy Investment
Financial commitments announced by Anthropic on the same day:
$200 million: establish the “Economic Future Research Fund” for piloting public policy
$150 million: carry out a nationwide grants program
At the end of the long-form piece, Amodei said: “The key challenge is not incentivizing growth, but finding ways for everyone to share in the benefits.” In the article, he confirmed that this framework “is the starting point, not the finish line.”
Comparison with Confirmed Positions of the Trump Administration
Confirmed position of the Trump administration: let the AI industry “grow wild,” beat China through positive competition, and proactively tear down state-by-state local regulatory barriers.
Amodei’s confirmed position: at the end of the long-form piece, seek bipartisan language: “These policy ideas have bipartisan common-sense appeal across the political spectrum; the sooner we act, the sooner everyone can share in the benefits of AI.” In the long-form piece, Amodei confirmed that in the past Anthropic had argued that “transparency requirements are enough,” and that this policy advocacy represents an official update to the organization’s position.
Frequently Asked Questions
What exactly are the mandatory testing quantitative thresholds proposed by Anthropic?
The 《Policy on the AI Exponential》 long-form article confirms three trigger thresholds; if any are met, mandatory independent audits must be completed before listing: training compute exceeding 10^25 FLOPs, annual AI revenue exceeding $500 million, or AI R&D spending exceeding $1 billion. The tests focus on four risk areas: cybersecurity, biological weapons, runaway AI systems, and automated R&D.
Why did Anthropic call for legislative regulation on the same day it released its strongest model?
Amodei’s confirmation logic is: Claude Mythos Preview can already find high-severity vulnerabilities in major operating systems, proving that the risk profile of AI is “clear enough to enable precise regulation.” Therefore, the earlier reason that the legislation was “not clear enough” no longer holds.
Has this policy proposal entered the U.S. legislative process yet?
As of June 11, 2026, 《Policy on the AI Exponential》 is a policy proposal document and has not entered any formal legislative process in the U.S. Congress. In the article, Amodei confirmed that the framework “is the starting point, not the finish line,” and no legislative proposals or specific implementation timelines have been released.
