After spooking Trump into safety testing, Anthropic AI models get global release

2026-07-01 · Source: AI - Ars Technica · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Intermediate, medium

Summary

Anthropic's Claude models, Fable 5 and Mythos 5, have received US export curb lifts, allowing global release for Fable 5 and restored US access for Mythos 5 since June 26. This follows a Trump administration directive that flagged the models as national security risks due to their advanced cybersecurity capabilities, particularly Mythos 5's ability to find and exploit software vulnerabilities. Anthropic agreed to expand government partnerships, establish a red-teaming program with hackers, and create a 24/7 internal team to monitor jailbreak threats. While Fable 5's safeguards are strengthened, addressing an Amazon-discovered bypass, this comes with a "tradeoff" of potentially blocking benign coding tasks. Anthropic is also collaborating with Amazon, Microsoft, and Google to draft a framework for assessing AI jailbreak severity.

Key takeaway

For AI security engineers evaluating frontier model deployments, Anthropic's experience highlights that government collaboration and proactive safety measures are critical for market access. You should prioritize establishing robust red-teaming programs and internal threat monitoring, while also preparing for potential user impact from tightened safeguards. Consider participating in industry efforts to standardize jailbreak severity assessments to streamline incident response.

Key insights

AI model export controls can be lifted through enhanced government collaboration and robust safety protocols.

Principles

• Frontier AI models require continuous red-teaming and 24/7 threat monitoring.
• Stronger AI safeguards may inadvertently block benign user tasks.
• Industry consensus frameworks are crucial for assessing AI jailbreak severity.

Method

Anthropic addressed export curbs by expanding government partnership, implementing a hacker red-teaming program, establishing a 24/7 internal jailbreak monitoring team, and developing an improved safety classifier.

In practice

• Implement a HackerOne program for security researchers to submit jailbreaks.
• Develop a safety classifier to block dangerous model behaviors.
• Route blocked requests to less powerful, safer models like Opus 4.8.

Topics

• AI Export Controls
• Anthropic Claude
• AI Safety Testing
• Model Red-Teaming
• Cybersecurity AI
• AI Jailbreaks
• Government-AI Collaboration

Best for: CTO, VP of Engineering/Data, Director of AI/ML, Policy Maker, AI Security Engineer, Tech Journalist

Related on AIssential

• Anthropic’s safety warnings may have just backfired — the government has pulled the plug on its most powerful AI — Government intervention in AI model deployment, spurred by developer safety warnings, highlights complex regulatory and commercial tensions.
• Anthropic's Amodei calls for workers’ rights over AI — Anthropic released Fable 5, a powerful AI model with robust guardrails, demonstrating a strategy for safe public deployment of advanced capabilities.
• Anthropic’s Safety Superpower — Anthropic's "safety" narrative strategically aligns with its economic and power imperatives to control frontier AI development and user data.
• Anthropic’s Fable 5 Is Coming Back. The 18-Day Standoff With Trump Is Over. — Government intervention in frontier AI model releases is now a de facto standard due to national security concerns over potential misuse and jailbreaks.
• The US Government Just Killed Two of Anthropic’s AI Models Overnight — The US government's first AI model recall signals a new era of export control for frontier systems.
• Anthropic comes to Washington to meet White House officials — The rapid advancement of frontier AI models is driving urgent calls for regulation and safety measures amidst geopolitical tensions and market expansion.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI - Ars Technica.