Anthropic’s AI Safety Concerns Lead to Government Restrictions on Its Most Advanced Model

Anthropic said Friday that it has been ordered by the U.S. government to immediately restrict access to two of its most advanced AI models, Claude Fable 5 and Claude Mythos 5, citing national security concerns. The company confirmed on X that it has complied with the directive, while also expressing disagreement with the decision.

According to Anthropic, the order was received at 5:21 p.m. ET on Friday and requires the company to shut down access to both models globally, not just for foreign nationals subject to the government’s export-control measures. The company noted that access to its other AI models remains unaffected.

The decision is significant because Mythos 5 is regarded as Anthropic’s most capable AI system. First previewed in April, the model has remained under tight restrictions due to what the company described as its exceptional ability to identify software vulnerabilities. Anthropic previously claimed that Mythos successfully detected flaws across major operating systems and web browsers during testing.

Rather than releasing the model publicly, Anthropic launched a controlled initiative known as Project Glasswing, allowing only about 50 vetted organisations to access the technology for defensive cybersecurity purposes. Participants reportedly included Amazon, Apple, Google, Microsoft, and CrowdStrike.

Fable 5, introduced earlier this week, was designed as a commercial version of Mythos with additional safeguards. Anthropic said the model included restrictions intended to block responses in high-risk areas such as cybersecurity and biology, making it suitable for broader public use. According to benchmark data from Vals AI, Fable 5 quickly became one of the most capable AI models available to general users.

Anthropic said the government’s action is formally being treated as an export-control measure, but the company believes concerns centre on reports of a possible jailbreak affecting Fable 5. According to Anthropic, officials provided only verbal evidence of what it described as a limited and non-universal exploit that involved prompting the model to analyse a codebase and identify software weaknesses.

The company argued that this level of capability is already available through several competing AI systems, including OpenAI’s GPT-5.5, and is routinely used by cybersecurity professionals for defensive security work.

Anthropic further stated that many of its most important safety protections operate independently of the model itself through external classifier systems. As a result, the company argues that even if users bypass certain conversational safeguards, additional protections remain in place to prevent the generation of the most dangerous outputs.

The company did not hide its frustration with the decision. In a public statement, Anthropic wrote that it disagrees with the idea that a narrow potential jailbreak should justify removing a commercial model already being used by hundreds of millions of people. The company warned that applying the same standard across the AI industry could significantly slow or halt future releases of frontier models.

The development comes as Anthropic is widely expected to pursue an initial public offering in the near future. The company has built much of its reputation around AI safety and responsible deployment practices. However, some observers have noted the irony that Anthropic’s own warnings about the potential risks of Mythos may have contributed to increased regulatory scrutiny.

Earlier this year, OpenAI CEO Sam Altman criticised Anthropic’s handling of Mythos, describing it during an interview with podcaster Ashlee Vance as “fear-based marketing.” Altman argued that portraying a model as exceptionally dangerous naturally attracts attention and concern. While he did not predict government intervention, critics now point to those comments as highlighting a challenge Anthropic faces today. When a company repeatedly emphasises the unique power and risks of its technology, regulators may take those warnings seriously as well.