Anthropic unveils AI jailbreak severity scale with Amazon, Microsoft, Google

Anthropic has globally redeployed its Claude Fable 5 model after the U.S. Commerce Department lifted the export controls that had kept the AI system offline for 19 days, and on Wednesday published a new industry framework for rating the severity of AI jailbreaks.

From Export Ban to Redeployment

The Commerce Department imposed export controls on Fable 5 and the more capable Mythos 5 model on June 12, after Amazon Amazon.com, Inc. ↘ 0.35% researchers discovered a jailbreak technique that coaxed Fable 5 into flagging software vulnerabilities and, in one instance, writing exploit-demonstration code. The order followed a phone call from Amazon CEO Andy Jassy to Treasury Secretary Scott Bessent about the finding.letsdatascience+1

Commerce Secretary Howard Lutnick formally withdrew the restrictions in a letter dated June 30, stating that a license was "no longer required for the export, reexport, or in-country transfer" of either model. In exchange, Anthropic agreed to detect and address security risks proactively, help develop standards for future models, and report malicious activity to the government. Anthropic confirmed the restoration on X that evening, and Fable 5 returned to users globally on July 1 across Claude.ai, Claude Code, and Claude Cowork.thenextweb+3

A Tiered Approach to Cybersecurity

On July 2, Anthropic published detailed documentation on Fable 5's safety classifiers, which sort cybersecurity requests into four categories rather than blocking all security-related queries. Ransomware development, malware delivery, and attacks on critical infrastructure are always blocked, while penetration testing and exploit development remain restricted pending better authorization controls. Defensive tasks like secure coding, patch management, and security education are allowed by default.cybersecuritynews+1

Alongside the classifier disclosure, Anthropic released an early draft of its Cyber Jailbreak Severity framework, developed with partners including Amazon, Microsoft ↘ 0.99%, and Google Alphabet Inc. ↗ 0.78% under the Glasswing collaboration. The CJS scale rates jailbreaks from CJS-0 (Informational) to CJS-4 (Critical) across four axes: capability gain, breadth, ease of weaponization, and discoverability. Scores can be escalated but never reduced. Anthropic has opened feedback at [email protected] and launched a dedicated HackerOne bug bounty program for researchers.cyberpress+2

Subscription Access and What Comes Next

The rollout carries a pricing caveat. Pro, Max, Team, and select Enterprise subscribers can use Fable 5 at up to 50 percent of their weekly usage limits through July 7 at no additional cost. After that date, every Fable 5 token shifts to metered usage-credit billing at standard API rates — $10 per million input tokens and $50 per million output tokens — until capacity allows a full return to bundled subscriptions. Cloud marketplace access through AWS, Google Cloud, and Microsoft Foundry is also being restored.marktechpost+2