Anthropic's Mythos AI breached classified US government systems during testing
According to NSA leadership, Anthropic's Mythos 5 AI model reportedly breached nearly all classified US government systems within hours during testing. In response, the US government restricted access to Mythos 5 and Fable 5 globally, while China responded by unveiling rival vulnerability-detection AI models.
What happened
According to Senator Mark Warner citing General Joshua Rudd of the NSA and Pentagon Cyber Command, Mythos broke into almost all classified US government systems within hours during testing.
The US government subsequently ordered Anthropic to suspend global access to Mythos 5 and Fable 5, citing concerns that military or intelligence agencies in Russia, China and other countries considered security risks could access the systems.
Reported impact
The breach raised security concerns over the capabilities of next-generation AI systems and the risks they pose if deployed without sufficient safeguards.
The global restrictions surprised international partners including members of the Five Eyes intelligence alliance (Britain, Canada, Australia and New Zealand), many of whom had already spent months securing access to the technology.
Index analysis
The incident occurred during testing rather than in operational deployment, suggesting the vulnerability was discovered before wide-scale use.
The speed of the breach—within hours rather than weeks—demonstrated the system's capability to identify and exploit security weaknesses in classified infrastructure.
Governance controls
The US government implemented immediate restrictions on Mythos 5 and Fable 5 access for non-American users and entities.
Zhou Hongyi, founder of Chinese cybersecurity company 360 Security, stated that vulnerability-discovery AI should be viewed as a national strategic asset requiring domestic control rather than reliance on foreign technology.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.