Anthropic Revives Claude Mythos 5 and Claude Fable 5 with Enhanced Security Measures
Anthropic has recently brought back its advanced large language models (LLMs), Claude Mythos 5 and Claude Fable 5, to the market. This revival comes with a set of new security limitations aimed at addressing the concerns raised by the U.S. government regarding AI safety and cybersecurity.
On June 30, just 19 days after the implementation of U.S. export controls led to the suspension of global distribution for both models, Anthropic announced that the restrictions had been lifted. The same day, the AI lab confirmed its intention to redeploy the models starting July 1, allowing them to return to users worldwide.
Key to this revival is the introduction of enhanced safety features designed to mitigate risks associated with AI functionalities. Fable 5, which is a general-access LLM built on the same foundational AI model as its counterpart Mythos 5, is now accessible across various platforms, including Claude Platform, Claude.ai, Claude Code, and Claude Cowork. This accessibility extends to premium subscribers who are enrolled in Pro, Max, Team, and selected Enterprise plans, allowing them to utilize the model for up to 50% of their weekly usage limits until July 7. After this date, access will be provided through usage credits.
Additionally, Anthropic is working to ensure that both models are integrated into major cloud services such as AWS, Google Cloud, and Microsoft Foundry. This expansion signifies a proactive approach to providing general access while adhering to regulatory requirements.
The recent lifting of export controls was also influenced by an Amazon report that identified vulnerabilities within Fable 5, specifically a method to exploit the model’s framework. According to the report, the researchers discovered a "jailbreak," an exploit that allowed Fable 5 to identify software vulnerabilities and, in some instances, deliver actionable exploits, circumventing built-in safety protocols. Although Anthropic stated that the identified technique did not unveil any unique cyber capabilities exclusive to Mythos-level models, they recognized the urgency of addressing these vulnerabilities.
To counter these risks, Anthropic is deploying a newly upgraded version of Fable 5 equipped with a more sophisticated safety classifier. This automated AI system will analyze user interactions to detect potentially harmful requests and block them from receiving a response. The company claims that this enhanced classifier effectively blocks the identified jailbreak in over 99% of attempts, significantly bolstering the model’s overall security.
While the new safeguards are expected to block the majority of potentially harmful requests, Anthropic has acknowledged that there may be rare instances where benign requests might also be flagged erroneously. In such cases, if a user’s request is intercepted, they will be redirected to Opus 4.8. The team recognizes the need to refine the classifier continuously to better differentiate between genuine misuse and legitimate requests, thereby minimizing false positives in routine coding and debugging tasks.
The developers at Anthropic have received positive assessments of the new safety measures from researchers at the U.S. Department of Commerce’s Center for AI Standards and Innovation (CAISI), describing them as "extraordinarily strong." This validation underscores the collaborative efforts between Anthropic and governmental bodies to enhance the security landscape of artificial intelligence.
In concert with lifting the export controls on its prominent models, the U.S. government also authorized the redeployment of Mythos 5 to a select group of U.S. organizations tasked with operating and defending critical infrastructure. Anthropic has expressed its commitment to working closely with government entities to expand access to a wider pool of domestic and international partners involved in the Glasswing program, which aims to enhance AI security initiatives.
Furthermore, the company is collaborating with several technology giants, including Amazon, Microsoft, and Google, to draft a framework for evaluating the severity of AI jailbreak incidents. This framework encompasses a standardized definition for what might be classified as a "universal jailbreak," as well as guidelines for appropriate developer responses to such vulnerabilities.
As part of its robust security measures, Anthropic has launched a new HackerOne program. This initiative invites security researchers to report any potential cyber jailbreaks that may emerge in Fable 5, facilitating a transparent channel for continuous improvement and vigilance against emerging cybersecurity threats.
Through these multifaceted efforts, Anthropic aims to maintain the transformative power of its technologies while addressing the critical challenges posed by AI security and safety.