Now loading...
The U.S. government has mandated an export control measure, causing Anthropic to suspend all access to its Fable 5 and Mythos 5 models for foreign nationals. This directive impacts all Anthropic customers, including foreign employees of the company, to ensure compliance with national security regulations. Access to other models offered by Anthropic remains unaffected.
The directive was received by Anthropic at 5:21 PM ET, although it did not elaborate on the specific national security concern prompting the suspension. According to the company, the government indicated awareness of a possible method for bypassing, or “jailbreaking,” Fable 5. A review by Anthropic revealed that the technique demonstrated identified minor vulnerabilities, which have been found in other publicly available models without requiring any form of bypass.
Anthropic reaffirmed its commitment to robust safeguards for Fable, which are designed to minimize the risk of misuse, particularly in cybersecurity. The company noted that some users have deemed these safeguards overly stringent. In preparation for Fable’s launch, Anthropic collaborated with the U.S. government, UK AISI, and various private organizations to rigorously test the model’s defenses over thousands of hours, confirming that Fable’s safeguards are more effective than previous iterations.
Despite extensive testing, no tester has succeeded in uncovering a universal jailbreak that could broadly bypass Fable’s security measures. Anthropic believes that achieving absolute jailbreak resistance may not be feasible for any model. There are potential vulnerabilities, termed non-universal jailbreaks, which could lead to minor cybersecurity information being elicited under specific conditions, and it’s likely that universal jailbreaks may eventually be discovered.
Anticipating these challenges, Anthropic implemented a “defense in depth” strategy for Fable 5 to ensure that any jailbreak attempts remain limited or costly, along with close monitoring to quickly respond to any successful breaches. The company has introduced a policy requiring the retention of customer data for 30 days, allowing for ongoing research into mitigating these jailbreaks, albeit with associated costs for customers.
Anthropic stands by its strategy, asserting that it effectively lowers the risks associated with Fable, aligning them with those of other models already in use throughout the industry. So far, the government has only cited verbal examples of potential non-universal jailbreaks, where the model was tasked with identifying and correcting specific code errors, reports of which are common in various existing models.
The company plans to share further information in the coming hours. While complying with the government’s directive to revoke access to Fable 5 and Mythos 5, Anthropic contests the basis for such a suspension, suggesting that a narrow finding of a potential jailbreak should not warrant a recall of a commercially deployed model serving millions of users. They argue that if a similar standard were enforced industry-wide, it might significantly impede new model innovations.
Anthropic has expressed its belief that the government should prevent unsafe deployments through a transparent, fair, and fact-based statutory process, which they feel has not been followed in this instance. The company apologized for the inconvenience caused to its customers and is actively working to restore access to its models.
