Anthropic's AI Model 'Mythos' Identified Vulnerabilities in U.S. Government Systems Within Hours, Report States

Anthropic's advanced artificial intelligence model, known as 'Mythos,' reportedly discovered vulnerabilities in highly secure U.S. government systems within hours during testing. The Associated Press reported this week that these tests, conducted by U.S. intelligence agencies, highlighted the AI's rapid ability to identify potential security weaknesses.

While the model identified vulnerabilities, the report clarified that this does not necessarily mean it could exploit them within the same timeframe. These tests were part of Anthropic's Project Glasswing, an initiative aimed at addressing potential security concerns surrounding its AI products. The company has faced scrutiny over the advanced capabilities of its models, including 'Mythos' and 'Fable,' with some warnings suggesting they could pose existential risks.

During a congressional hearing earlier this month, Sen. Mark Warner, the ranking Democrat on the Senate Intelligence Committee, stated he was informed by National Security Agency Director Gen. Joshua Rudd that Mythos had demonstrated the ability to breach "almost all of our classified systems, not in weeks, but in hours." A White House official confirmed that various government entities are employing cutting-edge AI models to pinpoint and mitigate cybersecurity vulnerabilities.

Earlier in June, the Trump administration imposed foreign export controls on the 'Mythos' and 'Fable' models. This action followed warnings from Amazon CEO Andy Jassy about evidence suggesting that the AI's safety guardrails could be bypassed. In response, Anthropic temporarily removed the models offline to comply with the new restrictions and dispatched officials to Washington, D.C., to engage with government representatives.

The National Security Agency had been evaluating versions of Anthropic's latest AI tools when these models were taken offline, underscoring the growing importance of these AI systems for cybersecurity and national security efforts.

Previously, tensions had surfaced between Anthropic and the U.S. government when CEO Dario Amodei declined to grant the Pentagon unrestricted access to AI tools during contract negotiations, citing concerns about their potential use in mass surveillance or weaponry. Following the imposition of export restrictions, Anthropic argued that the measures were unnecessary and that the company had merely flagged potential security risks within the AI.

However, White House officials reportedly expressed frustration that Anthropic had characterized these safety risks as a "narrow" problem, especially after years of the company highlighting the potential catastrophic consequences of uncontrolled AI. Anthropic is now reportedly seeking to collaborate more closely with the government to address security concerns, having recently proposed increased cooperation with the White House, according to reports. Talks between Anthropic and Trump administration officials are said to be progressing, though a definitive timeline for a resolution remains unclear.

At the G7 Summit in France, President Trump commented that discussions with Anthropic were proceeding well, while Amodei urged global leaders to maintain a unified approach to AI regulation rather than adopting fragmented strategies.