Training compute threshold
10²⁵ FLOPs
Revenue gate
$500M
R&D gate
$1B
Economic commitments
$350M
Anthropic asks governments to block dangerous AI deployments — and proposes a plan if AI displaces workers
On June 10, 2026, Anthropic published two policy frameworks it wants governments to act on: the Advanced AI Framework, which requests legal authority to block frontier model deployments that pose catastrophic risk, and the Economic Policy Framework, which maps US policy responses to AI-driven labor displacement across three unemployment scenarios. Both documents were released alongside $350 million in financial commitments and timed to follow the company's recursive self-improvement data from the same week.
Anthropic published two policy documents on June 10, 2026 that together amount to the most detailed regulatory ask the company has ever made: a framework for how governments should govern frontier AI models, and a separate framework for how the US should cushion the labor market if AI displaces workers at scale. 1 Neither is advisory background material — both request specific new legal authorities and concrete government spending — and both were timed to coincide with the company's recursive self-improvement report published the same week.
What the Advanced AI Framework asks for
The core request of the Advanced AI Framework is blunt: give governments the legal authority to block or deter AI deployments that pose a significant risk of catastrophic harm. 2 That authority does not exist in current US law, and Anthropic argues no proposals before Congress would create it.
The framework defines a narrow scope for when these obligations would apply: models trained using more than 10²⁵ floating-point operations, developed by companies earning more than $500 million in AI-related revenue or spending more than $1 billion on AI R&D. That threshold is calibrated to reach roughly the current frontier tier — it would cover Anthropic, OpenAI, Google DeepMind, and Meta, but not most smaller labs or open-source projects.
Four categories of catastrophic risk drive the framework:
- Biological. The same capabilities that accelerate drug discovery can lower the barrier for engineering dangerous pathogens. The framework cites this as the most near-term concern.
- Cyber. Claude Mythos Preview, per the document, already discovered thousands of high-severity vulnerabilities across every major operating system and browser. Deployed offensively, those capabilities could threaten critical infrastructure.
- Loss of control. As AI systems grow more capable, maintaining reliable control over their behavior may become harder — an argument connected directly to the company's published work on agentic misalignment.
- Automated R&D. AI systems now accelerate AI development itself, which Anthropic says could amplify the three risks above.
Loading stats card…
For frontier developers at that scale, the framework proposes four requirements:
Transparency — publish evaluation results, safety frameworks, system cards, and regular risk reports. California and New York already require some version of this; the Anthropic proposal goes further by mandating regular engagement with independent evaluators.
Independent evaluation — at least one qualified external evaluator should publish a review of each developer's evaluations and risk reports. Governments and industry would need to build out this ecosystem, setting standards and ensuring evaluators have sufficient model access.
Security — protect model weights and training infrastructure against both external attackers and internal threats. Developers should test their own defenses, report model distillation attacks to a designated agency, and disclose their security program publicly at a high level.
Enforcement with teeth — civil penalties tied to global annual revenue, escalating with repeated violations. The framework explicitly notes this goes beyond what current law or pending congressional proposals would provide. It also includes safeguards against regulatory overreach, and suggests policymakers could start with a lighter-touch mechanism and expand enforcement authority as the evaluation ecosystem matures.
The societal resilience section of the framework covers measures governments should take independent of what developers do: gene synthesis screening and biosurveillance for biological risks; hardening critical infrastructure software for cyber risks. On loss-of-control and automated R&D, Anthropic acknowledges the resilience agenda "is less developed" and flagged ongoing research rather than concrete proposals.
On the question of federal vs. state authority, Anthropic's position is distinctive: it explicitly says Congress should not preempt state law unless it enacts a federal framework at least as strong as what Anthropic is proposing. The company supported California's SB 1047 and similar laws, and that position hasn't changed.
What the Economic Policy Framework proposes
The second document takes a different tack. 3 Rather than telling governments what to do about AI capabilities, it asks what the US should do if those capabilities displace workers — and at what scale.
The framework describes three scenarios:
| Unemployment scenario | Primary policy response |
|---|---|
| ~5% (elevated but manageable) | Workforce training grants, licensing reform, wage insurance, pre-distributive capital accounts expanded to include AI equity stakes |
| ~10% (significant disruption) | Expanded unemployment insurance, sector-specific transition support, basic-needs relief; possible incentives to slow pace of AI deployment |
| Unprecedented levels | Income replacement at scale, new tax revenue sources, basic income or sovereign wealth models, equity-sharing mechanisms |
At the 5% level, the document focuses on making it easier for workers to retrain and enter new industries. One concrete proposal: expand the asset classes permitted in existing pre-distributive capital accounts — currently limited to index funds — to include direct stakes in AI companies, on the theory that workers should participate in the upside of the technology displacing them.
At 10% unemployment, the emphasis shifts from transition support to income stabilization. The framework also raises the possibility of incentivizing firms to manage displacement gradually, rather than rapid replacement of workforce capacity.
The unprecedented scenario is acknowledged as more speculative. Anthropic writes that "this scenario is novel economic territory, so we're less certain about the right answers here." The discussion of basic income and equity-sharing reflects that uncertainty rather than a firm proposal.
Two immediate financial commitments accompany the framework:
- $200 million to an Economic Futures Research Fund (an expansion of the Economic Futures Program launched a year ago) to fund policy research trials and program evaluations
- $150 million national fellowship program to help early-career workers extend AI benefits to communities across the US
The document notes that Anthropic is "willing to help fund" the responses it recommends — including ones not traditionally financed by private companies — and frames that as a form of accountability for the policies it advocates.
One framing choice stands out: the document opens with a statement that Anthropic is "not seeking job displacement" and is working to prevent or minimize it. The company adds: "Some amount of displacement, though we cannot say how much, may be an intrinsic consequence of the technology, and our responsibility is to prepare for it and respond to it." That is a narrower commitment than claiming AI development can be managed without workforce disruption, but it is also a more honest one.
Why both documents appeared together
Loading content card…
The timing connects both frameworks to the recursive self-improvement report Anthropic's Institute published on June 4, 2026. 4 That report showed Claude now authors more than 80% of Anthropic's production code and projected that AI task horizons are doubling roughly every four months — numbers that substantially shorten the timeline for when both sets of policy concerns become urgent.
Releasing the AI governance framework without an economic framework would have made Anthropic's regulatory ask look narrowly self-serving: focused on blocking competitors' deployments while ignoring the societal consequences of its own. Releasing the economic framework alone, without a governance mechanism, would have implied that workers can absorb the disruption on their own if the safety problem is handed off to governments. Publishing both together signals that Anthropic views these as parts of the same problem.
The policy documents are also explicitly addressed to governments outside the US. The Economic Policy Framework notes it is "focused on the US because we're an American company," but states the underlying principles are intended to be global. It mentions the upcoming AI Summit in Geneva and the G7 as forums where these proposals should appear. The AI governance framework similarly says that "addressing AI risks cannot wait for action in Washington."
What's missing
The frameworks are more detailed than most AI company policy documents, but several questions remain open.
On enforcement: who constitutes a "qualified independent evaluator"? The framework calls on governments and industry to build out that ecosystem, but does not specify what standards these evaluators should meet, how they would get meaningful model access, or what happens when evaluators disagree with a developer's own risk assessment.
On the economic side: the three unemployment scenarios are discrete thresholds, but the document doesn't specify what signals would indicate a transition from one scenario to the next, or who decides when the 10% regime should be activated. The framework acknowledges this uncertainty, but leaving it unresolved means the most consequential policy decisions would be made reactively.
The Advanced AI Framework's civil penalty mechanism includes unspecified "safeguards that would prevent that power from being misused" — but those safeguards are described as a design goal rather than a concrete proposal. For practitioners trying to evaluate regulatory risk, the absence of that detail is significant.
One open question the company did not raise: what happens to companies that currently sit below the 10²⁵ FLOP / $500M revenue threshold but are growing toward it? The framework's enforcement regime would only apply to the current frontier tier, potentially creating incentives for developers to maintain their capabilities just below the threshold.
Add more perspectives or context around this Post.