Anthropic Pulls Back Mythos-Class Models Globally

Brief Overview

• The US authorities have announced an export control order that impacts Anthropic’s Claude Fable 5 and Mythos 5 models.
• Access to these models is halted for all non-US individuals, including foreign Anthropic employees.
• The order is driven by worries about a potential jailbreak vulnerability in Claude Fable 5.
• Though Anthropic disagrees with the decision to recall, it is complying with the order to maintain security.
• Mythos 5 was launched under restricted conditions, featuring fewer protections compared to Claude Fable 5.

Effects of the US Export Control Order

In a notable action, the United States government has enacted an export control order mandating Anthropic to suspend access to its Claude Fable 5 and Mythos 5 models. This order influences all non-US nationals, including foreign employees of Anthropic, both in the US and abroad. The repercussions are extensive, impacting users of these models globally.

Security Issues and Adherence

Anthropic has stated that the action taken by the US government is premised on concerns regarding a possible jailbreak vulnerability in Claude Fable 5. The company characterized the vulnerability as “a potential narrow, non-universal jailbreak,” which entails directing the model to evaluate a specific codebase and correct software defects. Despite this, Anthropic has conveyed its disagreement with retracting a model already utilized by hundreds of millions worldwide.

Launch and Protections

Claude Fable 5, which became publicly available on June 10, was fitted with measures to mitigate misuse in the spheres of cybersecurity and biological and chemical research. In contrast, Mythos 5 was disseminated to chosen partners under Project Glasswing and lacked certain layers of protection present in Claude Fable 5, although both models utilize the same core technology.

Anthropic’s Reaction and Prospective View

Although not having received a formal notification of a non-universal potential jailbreak resulting in adverse effects, Anthropic has examined a report believed to underlie the US government’s prohibition. The company, in association with the US government and the UK’s AI Security Institute (AISI), conducted comprehensive evaluations of the models’ protections before releasing Claude Fable 5.

Anthropic upholds a 30-day data retention policy to track usage and has embraced a defence-in-depth approach to make jailbreaks difficult and expensive to achieve. Nonetheless, the company recognizes the challenges in achieving complete jailbreak resistance and expects to uncover minor circumventions of protections and potentially more extensive jailbreaks in the future.

Conclusion

The sudden suspension of Anthropic’s Claude Fable 5 and Mythos 5 models following a US export control order underscores the ongoing difficulties in AI model security and regulatory compliance. While Anthropic seeks to address the raised issues, the company continues to emphasize security and protective measures in its models.