JAKARTA - The United States government has ordered Anthropic to immediately stop access to its two most advanced artificial intelligence (AI) models, Claude Fable 5 and Claude Mythos 5, citing national security concerns.
Citing a report from TechCrunch, Saturday, June 13, Anthropic stated that it had complied with the order received on Friday at 17.21 Eastern Time. However, the AI company considered the government's decision to be inappropriate.
The order requires Anthropic to disable access to Fable 5 and Mythos 5 for all users in the world, not just foreign nationals who were previously the target of the United States' export control rules. Anthropic's other AI models can still be used.
Mythos is the most powerful AI model ever developed by Anthropic. The model was introduced in a limited way in April because it has a high ability to find security holes in software.
According to Anthropic's statement quoted by TechCrunch, Mythos was able to find vulnerabilities in various operating systems and major browsers that the company tested.
Because the ability is considered very sensitive, Anthropic does not open access widely. The company only allows about 50 selected organizations to use Mythos through a program called Project Glasswing, including Amazon, Apple, Google, Microsoft, and CrowdStrike for the sake of cybersecurity defense.
Meanwhile, Fable 5, which was released just a few days ago, is a Mythos version that has been equipped with guardrails or safety barriers. This technology is designed to prevent AI from providing answers in high-risk areas, such as cybersecurity and biology.
Based on the Vals AI test, the institution that monitors the performance of AI technology, Fable 5 is said to be one of the most advanced AI models available to the general public.
In the company's blog, which was also quoted by TechCrunch, Anthropic mentioned the government's concerns regarding the alleged jailbreak of Fable 5.
In the world of AI, jailbreaking is a technique for making an AI model bypass security rules or restrictions that developers have set.
However, Anthropic said the government has so far only presented oral evidence regarding the possibility of a limited jailbreak that does not apply to all conditions.
According to Anthropic, the ability to read program code and find software vulnerabilities is actually also available in other publicly accessible AI models, including OpenAI's GPT-5.5. This ability has also long been used by cybersecurity professionals to protect digital systems.
Anthropic also explained that its main protection system works through an independent classifier, namely a separate supervisory system that monitors and filters AI outputs. With this mechanism, protection against results considered harmful can still run even though there are attempts to bypass the initial rejection of the model.
However, the US government still carried out the decision. Anthropic expressed his disapproval of the move.
"We disagree that the finding of limited potential for jailbreaking should be a reason to withdraw a commercial model that has been used by hundreds of millions of people. If standards like this are applied across the industry, the launch of new AI models from all advanced AI developers could be hampered," wrote Anthropic.
The case also re-highlights a major debate in the AI industry: how far advanced AI capabilities should be restricted for security reasons, and how governments regulate a technology that is evolving so rapidly.
Previously, OpenAI CEO Sam Altman had criticized Anthrop's way of…
The English, Chinese, Japanese, Arabic, and French versions are automatically generated by the AI. So there may still be inaccuracies in translating, please always see Indonesian as our main language. (system supported by DigitalSiber.id)
