Anthropic Warns of Catastrophic Risks from New 'Claude Mythos' AI Model

Sundar Pichai in glasses on an orange background, beside a Google Discover interface with news cards for Cosan and Lincoln National, and rising graphs.

Key Facts

1Anthropic executives warned that the 'Claude Mythos' model poses catastrophic risks if released to the public.

2The model is capable of facilitating large-scale cyberattacks and terror attacks beyond current imagination.

The controversy surrounding the 'Claude Mythos' model has taken a dramatic turn following reports that the US military is utilizing the AI to intensify strikes in the ongoing conflict with Iran. Gregory Allen, senior adviser at the CSIS Wadhwani AI Center, confirmed the direct involvement of Anthropic's technology in these current military operations. Beyond the battlefield, the model is now being linked to the onset of a new era of AI-driven crime, heightening scrutiny from the Trump administration. This follows previous concerns regarding 'Project Glasswing' and the model's ability to identify critical vulnerabilities in global encryption standards like TLS and SSH. As Anthropic navigates this transition from safety concerns to military and criminal exploitation, it must balance its $800 billion valuation with the profound geopolitical implications of its technology.

about 2 months ago

Version History

Version 10about 2 months ago

What changed: The story was updated to include new warnings regarding the Mythos model's role in facilitating AI-driven crime alongside its military applications.

Version 9about 2 months ago

What changed: The story was updated to include reports of the US military using the Mythos model for combat operations against Iran, confirmed by CSIS advisors.

Version 8about 2 months ago

What changed: The story was updated to include White House intervention and escalating tensions with the Trump administration regarding the risks of the new model.

Version 7about 2 months ago

What changed: Updated with specific technical capabilities of the Mythos model, including CyberGym scores, the discovery of decades-old vulnerabilities in OpenBSD and FFmpeg, and the introduction of 'Project Glasswing'.

Version 6about 2 months ago

What changed: Anthropic officially launched the preview version of Claude Mythos on Wednesday, confirming it will remain restricted from the public due to its advanced capability to exploit software vulnerabilities beyond human skill levels.

Version 5about 2 months ago

What changed: The story was updated to include the model's discovery of critical software vulnerabilities and the formation of a major tech coalition to address these infrastructure risks.

Version 4about 2 months ago

What changed: Updated to clarify that Opus 4.7 serves as a strategic precursor and testing phase for safety protocols before the full Mythos rollout.

Version 3about 2 months ago

What changed: Updated the story to include the official rollout of Claude Opus 4.7 and clarified that it is intentionally less capable and risky than the Mythos Preview model.

Version 2about 2 months ago

What changed: Anthropic released an updated Opus model with improved engineering but weaker cyber skills than Mythos, amid a new $800 billion company valuation.

Version 1about 2 months ago

What changed: Updated the story to include Anthropic's admission that it can no longer fully measure the model's capabilities and is facing an internal oversight crisis.