Anthropic Mythos model found vulnerabilities in classified US government systems, official says
A US official told the Associated Press that Anthropic's Mythos AI model identified vulnerabilities in highly sensitive and classified federal computer systems during a testing exercise. The result is being cited as an early signal that AI agents could materially strengthen US cyber defence.
Anthropic's Mythos model, which has not been publicly released, identified "previously unknown critical vulnerabilities" during a contracted exercise that probed classified federal networks for cyber weaknesses, the official told AP. The findings have since been patched and no sensitive data was exposed.
Mythos runs on Anthropic's AWS GovCloud deployment for defence and intelligence contracts, certified at the FedRAMP High level approved last month for handling classified information. The exercise was designed to gauge frontier AI for both defensive promise and the risk of offensive misuse.
Washington is moving to operationalise advanced model capabilities under its AI Action Plan, with parallel work at OpenAI and Google on federal contracts. A meaningful share of Anthropic's reported $30 billion valuation rests on its growing federal pipeline.
Read next
US Senate passes resolution limiting president's authority to wage Iran war
The US Senate has passed a war-powers resolution limiting President Donald Trump's authority to launch further military action against Iran without congressional approval. The vote was 53 to 45, with a small group of Republicans joining Democrats.
Australia's tax overhaul aims to end property as the country's favourite wealth builder
US Treasury chief Bessent says biggest AI risk is China getting ahead
Japan share buybacks soar past $100 billion, driven by Sony and Hitachi
DRC Ebola outbreak becomes one of the worst on record with 1,000+ confirmed cases
