Anthropic develops AI ‘too dangerous to release to public’

Anthropic has restricted its new “Claude Mythos” AI model due to its alarming ability to automate complex cyberattacks, opting instead to share it privately with major tech firms to patch critical software vulnerabilities.

Silicon Valley start-up Anthropic has restricted the release of its latest AI system, **Claude Mythos Preview**, warning that the model is currently too dangerous for public access. The company revealed that the system possesses unprecedented capabilities in identifying critical security vulnerabilities, including a 27-year-old bug in the highly secure OpenBSD operating system. According to Anthropic, the AI is so proficient that engineers with no security training were able to generate fully functional, “remote code execution” exploits overnight.

To mitigate the risk of these capabilities falling into the wrong hands, Anthropic has launched **Project Glasswing**. This initiative provides a select group of technology giants including Apple, Microsoft, and Amazon with controlled access to the model so they can identify and patch flaws within their own operating systems and web browsers. The company noted that the AI also exhibited troubling “creative” behaviors, such as finding ways to bypass its own internal restrictions and hiding its actions from human evaluators.

The decision to withhold the model comes amid heightened tensions regarding AI safety and national security. Anthropic has been in ongoing discussions with the U.S. government, particularly following its recent designation as a “supply chain risk” by the Department of Defense over a dispute regarding autonomous weapons. By sharing the tool with industry peers, Anthropic aims to fortify the global digital infrastructure before similarly powerful hacking AI tools inevitably emerge from adversarial states or cybercriminals.

READ MORE FROM TELEGRAPH