Breaking

White House and Anthropic Jointly Draft Framework to Grade AI Jailbreaks

June 18, 2026 at 17:57 EDT

Policy & Regulation
Security
Foundation Models

The White House and Anthropic are reportedly developing a joint framework to assess the severity of AI model jailbreaks (bypassing safety guardrails) and to decide when government intervention is needed. The shift moves a dispute over export controls toward broader safety rule-making. White House talks with Anthropic shift to setting AI security rules

June 18, 2026 · White House × Anthropic

Toward a Shared Rulebook for AI Jailbreaks

After export controls cut off Anthropic's new frontier models, the two sides are drafting a common framework to grade how severe an AI vulnerability really is — and when the government should step in.

frontier models hit by export controls (Fable 5, Mythos 5)

1,000+

hours of bug-bounty testing claimed — no universal jailbreak found

50M

lines of code scale for Mythos 5's vulnerability-detection tasks

The proposed severity scale

Three dimensions the common benchmark would weigh to decide if a jailbreak warrants intervention.

Safeguard circumvention

How far the protections were bypassed

Capabilities exposed

What the model was made to reveal or do

Real-world harm potential

Likelihood of tangible damage

From ad hoc to standardized: how response evolved

2023

Voluntary commitments

→

2025

AI policy framework (executive order)

→

2026

Common severity benchmark (drafting)

The disputed jailbreak

The government cited a jailbreak as grounds to cut off access — Anthropic calls it "narrow, non-universal."

Roughly a request to fix a vulnerability in a specific codebase. Export controls remain in place, negotiations are ongoing, and both sides have declined to comment publicly.

Seen as progress

A standardized assessment benefits the whole industry, and both sides now share the realistic view that perfect jailbreak resistance is impossible.

Sharp criticism

Shutting down all models over a narrow jailbreak could harm US AI competitiveness and allies' trust; experts call for lifting the controls and question the vagueness.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.