Anthropic’s most sophisticated public AI model was suddenly stripped from customers within days of release.
The Trump administration called the model a national security risk after being notified of jailbreaks and ways to circumvent domestic guardrails. An export ban was placed on this model, which prevented some Anthropic employees from using it.
But Anthropic says the vulnerability doesn’t justify such an extreme reaction.
A number of perspectives highlight the confusing state of AI regulation in the United States. Experts say governments should be involved in discussions about AI safety, especially those that affect national security. But the recent spat between Anthropic and the government has brought broader concerns to the surface. There is no transparent and consistent framework for regulating AI, which could suffocate the U.S. industry.
And Anthropic, a high-profile AI research lab valued at nearly $1 trillion that is about to go public, is at the center of these tensions.
The government’s move comes after the AI company had a disagreement with the Pentagon over requests for changes to guardrails for military AI systems, leading to the Pentagon blacklisting Anthropic as a “supply chain risk.” Cybersecurity concerns then became widespread as the company said its latest AI model, Mythos, was highly adept at finding security flaws. Anthropic only released the full version of Mythos to a select group of partners before releasing a public version with guardrails called Fable 5 on June 9th.
Now, the Trump administration says some of those guardrails could fail, allowing hackers to add highly capable tools to their arsenal. Earlier this month, Anthropic suspended all access to both Mythos and Fable in order to comply with government directives.
Some experts are therefore calling for a clearer window into how governments make cybersecurity decisions. Otherwise, the United States risks falling behind in the global AI arms race.
“The problem is not that the government exercised discretion. National security requires such discretion,” Jessica Tillipman, associate dean of government acquisition law at George Washington University, wrote in an essay. “What’s surprising is that there’s no meaningful process at all.”
Anthropic and the administration are meeting to resolve the issue. President Trump said at the G7 summit on Wednesday that negotiations with Anthropic are “progressing well.” President Trump told Axios in an interview published Friday that he no longer considers the company a national security threat. “Well, maybe not now, maybe a week ago.”
The White House referred CNN to the Commerce Department, which did not respond to a request for comment.
Current state of AI regulation
The Trump administration has so far taken a light approach to regulating artificial intelligence, hoping to accelerate advances in the technology to keep the United States ahead of adversaries such as China.
The administration has reversed Biden-era policies such as mandatory safety reporting standards in favor of a voluntary framework and preemption of state laws. In March, President Trump announced the National Policy Framework on AI, which serves as a blueprint and set of recommendations for Congress to develop AI policy. However, the proposed plan says Congress should not regulate AI through a single rule-making body, but rather through sector-specific regulatory bodies. It also says national security agencies need to understand frontier AI models and their potential security risks.
Earlier this month, the administration issued an executive order requiring AI companies to voluntarily share their latest models with the government for cybersecurity review before making them public. But that order was also delayed at the last minute after President Trump said he was concerned it would “get in the way” of U.S. AI innovation.
Some states have passed their own laws and sued AI companies over safety concerns. California, for example, passed a law requiring AI companies to publish risk frameworks, report safety issues, and provide whistleblower protection. The state of Florida has launched a criminal investigation into OpenAI and is suing the company. The state alleges that OpenAI harms children and consumers, and that ChatGPT may have aided and abetted last year’s mass shooting at Florida State University. OpenAI denies these claims and says it takes safety seriously through various guardrails and initiatives.
Brad Carson, director of Public First, a bipartisan pro-AI safety super PAC, said the government should help establish AI safety protocols in a consistent, fair and clear manner.
“Right now, we have an ad hoc, personalized, opaque and possibly lawless approach,” he said.
The company said in a June 12 statement that the government did not provide Anthropic with specific details about the national security concerns that led to the export restrictions for Fable 5 and Mythos 5. Anthropic was initially given just 90 minutes to complete the model, a source close to the company told CNN.
On Monday, dozens of cybersecurity researchers, AI entrepreneurs, and business executives signed an open letter criticizing the government’s actions and calling on the Trump administration to engage in “an open, scientific, and transparent process to address AI risk assessment in the future.”
Additionally, he noted that advanced AI models can also be used by well-intentioned actors.
“It is dangerous to extract the best from the defenders without good reason when the enemy is rapidly advancing,” they wrote in the letter.
Some people question the significance of the jailbreak. Amazon first reported the matter to the government, a person familiar with the matter said. Anthropic said the vulnerability is minor and may exist in other companies’ models as well.
Several AI safety researchers, including former Facebook chief security officer Alex Stamos, have reviewed the research behind the government’s decision and said they disagree with that assessment.
“While there were some valid findings, there were no unique features that warranted a response anywhere close to this,” Stamos wrote in X.
David Sachs, a Trump advisor and former White House AI czar, dismissed the idea that the jailbreak was anything serious.
“It is difficult to see how they can argue that a jailbreak that allows the operation of cyberweapons can be defined as anything other than ‘serious,'” Sachs wrote to X.
Some worry that the administration’s actions against Anthropic could set a broader precedent.
“The harm goes beyond just one company. An administration governing in this manner will not be able to avoid the harsh regulations we fear,” Tillipman wrote. “In every cycle I’ve documented, it’s creating the conditions for catastrophe and abuse, and it’s causing exactly that reaction.”
