Anthropic’s powerful “Mythos” AI model was reportedly able to find vulnerabilities in highly secure US government systems within just a few hours — the latest development to stoke security concerns about the product.
During tests conducted by US intelligence agencies, Mythos identified certain vulnerabilities in government systems within hours – though that does not mean the bot would be able to exploit those same sensitivities within that timeframe, the Associated Press reported this week.
Those tests were part of an Anthropic initiative known as Project Glasswing, which partnered with other companies and agencies to resolve potential security concerns – following reports Anthropic’s Mythos and “Fable” were so advanced that they could spark an AI doomsday.
The dire warnings about the software have prompted skepticism from some observers, who say they’re basically marketing for Anthropic in the highly competitive AI race.
But at a congressional hearing earlier this month, Sen. Mark Warner, the top Dem on the Senate Intelligence Committee, said he’d been informed by National Security Agency Director Gen. Joshua Rudd that Mythos “broke into almost all of our classified systems, not in weeks, but in hours.”
A White House official told The Post that various parts of the US government have been using cutting-edge AI models to identify and mitigate cyber vulnerabilities, and that President Trump has taken action to protect classified systems.
The National Security Agency and Anthropic did not immediately respond to The Post’s requests for comment.
Earlier this month, the Trump administration slapped foreign export controls on “Mythos” and “Fable,” after Amazon CEO Andy Jassy warned the administration that researchers had found evidence it was possible to bypass their safety guardrails.
Anthropic responded by pulling the models offline entirely, claiming that was the only way to comply with the new restrictions – as it sent several top officials to Washington, DC, in an attempt to win over government officials.
The National Security Agency had been testing versions of Anthropic’s latest AI tools when the latest models were shut down, according to the New York Times, in a sign of how important the bots have already become for cybersecurity and national security measures.
Meanwhile, tensions between Anthropic and the US government have been heating up ever since Anthropic CEO Dario Amodei refused to give the Pentagon unchecked access to AI tools during contract negotiations, seeking limits for their use in mass surveillance or weaponry.
After the government slapped Anthropic’s latest models with export restrictions, the company argued the steps were unnecessary, saying it had simply flagged potential security risks in the bots.
But as The Post previously reported, White House officials were irked that Anthropic had downplayed the safety risks as a “narrow” problem – after years of warning about the potential catastrophic consequences of unruly AI bots.
Now Anthropic is scrambling to cozy up to the government and resolve security concerns, most recently pledging to work more closely with the White House in a proposal to Commerce Secretary Howard Lutnick, as The Post exclusively reported.
Talks between Anthropic and Trump officials are progressing well, though a timeline for a permanent fix remains unclear, a source said.
At the G7 Summit in France last week, Trump said talks with Anthropic were “going fine,” while Amodei urged world leaders to “resist the temptation to splinter” in their approaches to AI regulation.