US artificial intelligence firm Anthropic has developed a new AI model, Claude Mythos Preview, which it describes as too powerful for public release. The company is currently in discussions with the US government regarding the model’s capabilities and potential risks.
According to Anthropic executives, Claude Mythos Preview demonstrates an unprecedented level of autonomy, capable of reasoning like an advanced security researcher. The model can identify tens of thousands of software vulnerabilities and, unlike previous iterations, generate corresponding exploits.
Anthropic’s head of frontier red team, Logan Graham, characterized the development as “the starting point for what we think will be an industry change point, or reckoning, with what needs to happen now.” The company has stated that the model will be made available only to a select group of technology and cybersecurity firms through a new industry partnership called Project Glasswing.
Project Glasswing includes over 40 organizations, among them Apple, Amazon, Microsoft, Google, and NVIDIA. The initiative aims to test the model’s ability to identify and help remediate vulnerabilities in critical software systems. Anthropic has emphasized that public access will not be granted until appropriate safeguards are implemented to limit the model’s most potentially dangerous capabilities.
The development comes amid heightened scrutiny of AI security. Last month, US President Donald Trump issued a six-month ban on government agencies using Anthropic’s AI, citing concerns about the company’s influence on Pentagon operations and national security. The Department of Defense subsequently entered into an agreement with OpenAI for classified military applications.
Anthropic has acknowledged inadvertently leaking internal materials related to Claude Mythos in February when thousands of documents were left in a public data cache. Earlier this month, the company accidentally published over 500,000 lines of secret code, including unreleased features and developer notes, which it attributed to “human error, not a security breach.”
The company has provided the model to external groups, including US government organizations, to assess key risks such as cybersecurity vulnerabilities, loss of control scenarios, chemical-biological-radiological-nuclear threats, and harmful manipulation potential. These assessments are being incorporated into Anthropic’s overall risk evaluation framework.
The emergence of Claude Mythos Preview highlights the ongoing tension between AI advancement and security concerns, as companies navigate the challenges of developing increasingly capable systems while managing potential risks to critical infrastructure and national security.
