Inside the White House–Anthropic AI Safety Framework: Grading Jailbreaks, Export Controls, and the Push for Rules-Based Oversight

A Dispute Over What Constitutes a Dangerous Jailbreak

The immediate trigger for the export controls was a reported jailbreak — a method to bypass the models' safety guardrails. But the two sides disagreed sharply on its severity.

• Anthropic's position: CEO Dario Amodei argued during emergency calls with officials that the jailbreak in question was "narrow" and "non-universal" — not a systemic bypass that could be broadly exploited. The company stated publicly that the government had only provided "verbal evidence of a potential narrow, non-universal jailbreak" . “We disagreed that the finding of a narrow potential jailbreak should be cause for recalling a commercial model,” Anthropic said in a blog post .
• Government's position: The White House viewed any demonstrated bypass of safety guardrails on a model with advanced cyber-offense capabilities — Fable 5 was designed in part to identify software vulnerabilities — as an unacceptable national security risk, particularly if the models remained accessible to foreign adversaries .

The disagreement escalated rapidly. According to multiple reports, Amodei pushed back forcefully in calls with administration officials, describing the government's concerns as a misunderstanding, and at one point negotiations nearly broke down entirely .

The White House designated Fable 5 as a test case for its new AI executive order guardrails, coordinating across Commerce, Treasury, and the White House itself . The administration feared the models' capabilities could enable Chinese access to cybersecurity offense tools and other dangerous dual-use technologies .

From Confrontation to Negotiation: Building a Unified Compliance Framework

Within days of the export control crisis, the White House and Anthropic shifted from confrontation to structured negotiation . On April 17, 2026, Amodei met with White House officials in what were described as productive truce talks — the first senior-level contact since earlier friction between Anthropic and the Pentagon .

Key figures in the subsequent high-level discussions include U.S. Secretary of Commerce Howard Lutnick, National Cyber Director Sean Cairncross, and Anthropic co-founder Tom Brown . The administration has coordinated across Commerce, Treasury, and the White House for the Fable 5 test case .

The central project to emerge from these talks is a unified compliance framework for evaluating the severity of security vulnerabilities — particularly jailbreaks — in frontier AI models . The framework is designed to create common benchmarks for assessing whether a safety flaw justifies government intervention .

The Unified Grading System: Three Criteria

Central to the effort is a unified grading system for jailbreaks. While no official labels (e.g., "Critical," "High," "Medium," "Low") or a detailed rubric have been publicly released, the framework reportedly assesses vulnerabilities along three risk-assessment criteria :

1. Bypassed protections — Which safety guardrails (e.g., constitutional classifiers, real-time filters) were defeated.
2. Exposed capabilities — What dangerous model capabilities (e.g., cyber offense, biological guidance, chemical synthesis) became accessible.
3. Real-world impact — How easily the jailbreak could be exploited at scale and what harm could result.

The government's goal is to use this grading system to establish clear, pre-defined risk thresholds that would automatically determine whether to impose export controls, require mitigation, or block deployment — rather than acting ad hoc after a crisis .

The Government's Strategic Goal: Moving from Reactive to Rules-Based

The overarching government objective is to move from reactive, crisis-driven intervention to a rules-based system . By co-developing the grading framework and risk thresholds with Anthropic, the White House aims to:

• Create predictability for AI labs about what triggers government action.
• Establish shared terminology for describing vulnerability severity.
• Avoid future scenarios where a model is publicly deployed, then abruptly shut down under export controls — as happened with Fable 5 and Mythos 5.

Key uncertainty: No detailed criteria, scope, or timeline for the framework have been publicly released. The three risk-assessment criteria (bypassed protections, exposed capabilities, real-world impact) have been reported by multiple outlets but the specific scoring methodology has not .

The Broader G7 Context

The U.S.-Anthropic framework development is taking place against the backdrop of G7 discussions on AI safety standards. The G7 has been working toward harmonized international benchmarks for evaluating frontier AI risks . The unified grading system the White House and Anthropic are building is expected to inform — and be informed by — those broader multilateral talks, though specific G7 outcomes remain in negotiation .

What Comes Next

The framework's development marks a significant pivot in U.S. AI policy. Just days before the export controls, Amodei had published a policy essay arguing that governments should hold mandatory legal authority to block or reverse the deployment of frontier AI models that fail independent safety testing . The June 12 crisis put that principle to an unexpectedly confrontational test.

The key question that remains open: how to define the line between a narrow, manageable vulnerability and a systemic risk that warrants government intervention. The answer will shape not only Anthropic's future releases but the broader regulatory landscape for frontier AI development.