AP said on the 23rd that Anthropic's artificial intelligence (AI) model "Mythos" found vulnerabilities in a U.S. government classified system in just a few hours. Mythos is Anthropic's advanced AI model with expert-level cybersecurity capabilities.
According to AP, in a recent test that Anthropic conducted in cooperation with U.S. intelligence agencies, Mythos found weaknesses in a tightly secured government computer system. Sen. Mark Warner of Virginia mentioned the test at a congressional hearing this month, saying, "Mythos penetrated virtually almost all classified systems in hours, not weeks."
However, although Mythos discovered vulnerabilities, it did not appear to have gone so far as to exploit them.
Anthropic recently released the enterprise "Mythos 5" and the safety-enhanced "Fable 5" models to the public, but the U.S. government issued export restrictions blocking foreign access to the two models, citing national security concerns. As a result, Anthropic has suspended "Fable 5" and "Mythos 5" services for all customers.
More than 100 cybersecurity officers and experts from major U.S. tech corporations such as Adobe and Nvidia also sent a letter asking for the withdrawal of the government's control guidelines, saying they could instead benefit adversarial nations.