← The Polylog AI Intelligence Brief
Morning Edition · Sunday, June 21, 2026Published at 6:59 AM EDT · New York
Amodei Asks Government for the Power to Block Unsafe Models
Anthropic's "Policy on the AI Exponential" proposes binding rules tied to a compute threshold, a notable step from a frontier lab that is still releasing frontier models.
Anthropic has published Policy on the AI Exponential, a pair of proposed regulatory frameworks accompanied by an essay from chief executive Dario Amodei. The central request is unusually direct for an industry that has resisted binding rules. Anthropic wants government to hold the authority to block or deter dangerous deployments, a structure that VentureBeat compared to aviation safety regulation.
The Advanced AI Framework would apply only above defined thresholds, models trained on more than 10^25 floating-point operations (FLOP), built by companies earning more than 500 million dollars in AI revenue or spending more than 1 billion dollars on research and development. Covered models would face mandatory third-party testing across four risk areas, cybersecurity, biological weapons, loss of control, and automated research that could compound the other three. A companion Economic Policy Framework outlines a tiered response to AI-driven job loss, backed by a 350 million dollar pledge.
The timing invites scrutiny. Anthropic is calling for binding limits days after a government order pulled two of its own models offline, and a compute-and-revenue threshold conveniently draws the regulatory boundary around a small set of incumbents. The research behind these claims comes from the company's Societal Impacts and economics teams, not from independent evaluation.
The proposal is best read as a view of where governance is heading, not as settled policy. It concedes that voluntary transparency is no longer sufficient, which itself shows how far the regulatory conversation has moved.
What this means
A leading lab arguing for government authority to block its own product category signals that the industry now expects binding regulation and would rather shape the threshold than oppose the principle. A 10^25 FLOP threshold set at the frontier also functions as a barrier to entry, raising compliance costs that weigh most heavily on smaller entrants and open-weight projects.
What to watch
Part of a tracked trend
Oversight and Evaluation Lag Accelerating AI Capabilities
Over the next 3-6 months, evidence mounts that governance, evaluation, and agent-safety methods are failing to keep pace with capability growth, driving investment in interpretability, agent-manipulation benchmarks, and institutional-reform proposals.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Washington Turns Export Controls on a Model Itself, Forcing Anthropic to Disable Fable 5 and Mythos 5
- The NSA, Mythos, and the Quiet Arrival of an AI Cyber Doctrine
- Nvidia Raises 25 Billion Dollars in Bonds to Fund the AI Buildout
- Meta's SAM 3 Pushes Open-Vocabulary Segmentation Toward Production
- Genesis AI Unveils Eno, a Wheeled Robot Built on a Robotics Foundation Model
- Meta's Muse Spark Marks a Break From Llama, and From Open Weights
- OpenAI Hires Trump AI Plan's Architect to Run a New Strategic Futures Team
- OpenAI Adds Record-and-Replay Automation to Codex on macOS
Comments
0No comments yet.