The US government has ordered the suspension of access to advanced artificial intelligence (AI) models Claude Fable 5 and Mythos 5 for all foreign nationals within the country and abroad, including foreign employees of the company Anthropic, which produces them. As the firm said in a statement, the reason for the decision was cybersecurity problems that, according to authorities, threaten national interests.
Anthropic noted that it ended up temporarily disabling Claude Fable 5 and Mythos 5 for all customers to comply with the requirements. According to the corporation, the government provided “only verbal evidence” of a potential, highly targeted and uncommon hacking method that is accomplished by asking the model to check its underlying code for errors.
The development company reviewed a demonstration of this particular method, which was used to identify a small number of previously known minor vulnerabilities. All of them are “relatively simple,” according to Anthropic, and other “publicly available models” can also detect them “without the need to bypass the system.”
“We do not agree that the discovery of a potential security loophole should trigger a recall of a commercial model that supports hundreds of millions of people,” the company said in a statement.
Anthropic clarified that not a single supplier of AI models can guarantee the complete absence of so-called jailbreaks – bypassing the models’ protective mechanisms.
Moreover, with the release of Fable 5 in early June, the developers said that universal jailbreaks might be found in the future. At the same time, in the weeks leading up to the launch of the model, joint vulnerability tests were conducted with the US government, the UK Cyber Security Agency, several private third-party organizations and internal teams.
These tests, according to company representatives, showed that the Fable's security features are significantly more effective than any of the previously used models. During the tests, not a single tester was able to find a universal way to bypass the protection.
“We have implemented robust security measures that significantly reduce the likelihood of Fable being misused to solve cybersecurity problems. In fact, our protection measures are so strong that many users complain that they are too strict,” Anthropic emphasized.
The company apologized to customers for the incident and announced that it was working to restore access. Although both models were built on the same technical foundation, only Fable 5 was released to the public. It places much stricter restrictions on the types of questions users can ask, especially questions related to cybersecurity and biology, NBC News points out.
Mythos 5, in turn, was presented without such restrictions to a select group of trusted partners, including key cybersecurity and infrastructure companies, the channel adds. As RTVI.US wrote earlier, the capabilities of Mythos seriously frightened not only the United States, but also China. AI is capable of hacking government databases, financial systems and medical networks, identifying their vulnerabilities.
Anthropic's decision to suspend user access to its platforms marked the first time a leading artificial intelligence company has disabled its public model due to federal government intervention, NBC News said.
The actions of the American authorities regarding Anthropic indicate an increasing desire in Washington to consider advanced artificial intelligence systems as a tool for ensuring national security, writes Axios. According to the publication, US Secretary of Commerce Howard Lutnick sent a letter to the company's CEO Dario Amodei, in which he said that the models would be subject to export controls.
This is the second time that the Trump administration has taken action against this technology company, writes The New York Times (NYT). The newspaper mentioned “intense negotiations” in March this year about how Anthropic's artificial intelligence could be used in the military and intelligence sectors.
The Pentagon said then that it considers the corporation an “unacceptable source of risk,” which does not allow federal agencies to use its tools, the publication notes.
