Anthropic Withholds Advanced AI Model Claude Mythos Due to Cybersecurity Risks

Anthropic announced on Tuesday its new AI model, Claude Mythos Preview, capable of identifying thousands of previously undetected software vulnerabilities in common applications, operating systems, and browsers.

The model has exposed exploits for which no patches exist. Anthropic determined that widespread release without safeguards would enable cyberattacks, leading to restricted distribution. The company formed an alliance with cybersecurity firms and major tech companies to use Mythos for scanning and securing software.

Anthropic is exploring designing its own chips to support advanced AI development. Access remains limited to these partners to prevent misuse.

Response US Treasury and Federal Reserve officials convened Wall Street leaders to address potential systemic threats from the model. Bloomberg reported the meeting focused on potential systemic threats from the model's hacking capabilities. No public details emerged on the meeting's outcomes.

Cybersecurity experts have previously warned that advanced AI models enhance hacking speed and scale, allowing attacks comparable to state-sponsored operations. Anthropic's prior model, Claude Opus 4.6, was less capable at autonomously finding exploits.

Mythos represents a shift.

Anthropic, a San Francisco-based AI startup, positions Mythos as a tool to bolster defenses against hacking. The model identifies weaknesses at a sophistication level previously limited to elite groups in countries like China, Russia, and the United States.

Instances of AI models from Anthropic, OpenAI, and Google being used in cyberattacks by criminal and state-backed groups have been reported. By withholding public release, Anthropic aims to mitigate risks to critical infrastructure and government services worldwide.