Anthropic's Mythos AI model found vulnerabilities in classified US government systems

Anthropic's Mythos AI model has successfully identified vulnerabilities in classified US government systems, a US official confirmed to the Associated Press — marking one of the most consequential public disclosures of AI being used on sensitive government infrastructure to date.

The finding underscores a dramatic shift in how frontier AI models are being deployed. Rather than serving purely as productivity or research tools, systems like Mythos are now being directed at America's most sensitive networks, tasked with finding the security gaps that human analysts or conventional tools might miss. That an AI identified real vulnerabilities in classified environments suggests the technology has crossed a meaningful threshold in both capability and trust.

The disclosure comes at a turbulent moment for Anthropic. The company is simultaneously at the centre of a separate legal dispute over US export restrictions that have blocked customers from accessing its advanced AI models abroad. Together, the two developments paint a picture of a company whose technology is deeply embedded in US national security considerations — both as a tool the government wants to use and as an asset it wants to control.

For the broader AI industry, the Mythos episode is likely to accelerate debate about governance: who authorises AI to probe classified systems, who is liable if something goes wrong, and whether adversarial nations can reverse-engineer capabilities from such disclosures. India's own nascent AI security programmes will be watching closely.