Trump Restricts Anthropic AI Models After Security Breach Fears

Anthropic’s powerful AI model, Mythos, reportedly penetrated classified U.S. government systems within hours, triggering major alarm among American intelligence officials. The report, published overnight between Wednesday and Thursday by AP, said the model uncovered critical security weaknesses during testing by U.S. intelligence agencies as part of Anthropic’s Glasswing project.

At a congressional hearing earlier this month, Senator Mark Warner, a senior member of the Senate Intelligence Committee, cited National Security Agency Director Gen. Joshua Raud as saying, “Mythos penetrated almost all of our classified systems, not in weeks, but in hours.” That warning helped fuel growing concerns over the model’s capabilities and the risks it posed to sensitive systems.

In response, the Trump administration imposed export restrictions on Anthropic’s Mythos and Fable models after Amazon CEO Andy Jassy warned officials that researchers had found ways to bypass the models’ safety mechanisms. Anthropic then removed the models from the internet, saying that was the only way to comply with the new restrictions.

The company also sent senior executives to Washington in an effort to persuade officials, while White House sources expressed anger that Anthropic had downplayed the seriousness of the risks. Anthropic is now trying to align itself with the administration, and talks between company representatives and Trump administration officials are reportedly progressing well, though no final timeline has been set. Trump said during last week’s G7 summit in France that the discussions were “going well,” and Anthropic CEO Dario Amodei urged world leaders to “resist the temptation to split apart” in their approach to AI regulation.