What is the framework about?
In June 2026, Anthropic released its Advanced AI Framework, a set of policy proposals describing what the company believes governments should do in the near term about the most serious risks from frontier AI.
The framework is written primarily with the U.S. federal government in mind, though Anthropic suggests many underlying principles apply more broadly and encourages policymakers in other jurisdictions to tailor the recommendations to their own capacity and authority (p. 1).
Anthropic frames the framework as a starting point rather than a finished design, stating it is more confident about some parts than others (p. 2).
1: What obligations are proposed for AI frontier developers?
The first part proposes obligations for companies building the most capable models. Under these proposals, developers would test their models for a defined set of catastrophic risks, demonstrate mitigations, publish their findings, and answer to a government agency (p. 3).
The obligations are scoped to developers meeting both a technical threshold (models requiring more than 10²⁵ training FLOP) and an economic threshold (more than $500 million in annual AI-derived revenue, or more than $1 billion in annual AI R&D spending) (p. 4).
The framework prioritizes four "Enumerated Risk" categories: biological weapons, offensive cyber operations, loss of control, and automated research and development (p. 4).
On transparency, Anthropic recommends that covered developers publish a safety framework, issue a risk report at least every six months, release system cards when deploying materially more capable models, and report critical safety incidents to the designated agency within 15 days of discovery (p. 5–p. 7).
The framework also recommends independent evaluation, proposing that developers engage at least one qualified, financially independent evaluator within six months of the regulation's enactment (p. 7).
Additional sections address security programs for model weights and infrastructure (p. 9), and enforcement mechanisms including civil penalties, prohibitions on false statements, whistleblower protections, and possible authority for an agency to block or deter deployment of models posing significant catastrophic risk, paired with safeguards against regulatory overreach such as court enforcement and judicial review (p. 10–p. 12).
2: What are societal resilience measures as AI advances?
The second part shifts from developer conduct to how society can withstand threats that advancing AI may accelerate. Anthropic notes that the most consequential resilience investments take years to build and cannot be assembled during a crisis, so the recommendations focus on building them now (p. 13).
The biological resilience measures are organized as layers of defense—prevention, detection, and preparedness—including modernized biosafety standards, gene synthesis screening, pathogen-agnostic biosurveillance, and stockpiled protective equipment (p. 13–p. 16).
The cyber resilience measures aim to ensure that the same technology lowering the cost of attack also lowers the cost of defense, covering open-source and legacy software security, capability tracking, defender-advantage tooling, and security modernization (p. 16–p. 18).
Anthropic acknowledges that the societal resilience agenda for loss of control and automated R&D is less mature than for biological and cyber risks and states it will share more as its understanding develops (p. 19).
3: Closing
Anthropic characterizes these measures as necessary first steps, argues that waiting for perfect policy carries its own cost during a critical period, and invites feedback from policymakers and experts (p. 19).
Source: Anthropic. (2026, June). Anthropic's Advanced AI Framework. https://www-cdn.anthropic.com/files/4zrzovbb/website/0a58d567024a8b448ff15158ebc3625328dfcc1f.pdf
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.