Anthropic limits rollout on new cybersecurity model

Summary

• Anthropic is restricting public access to its newest advanced AI model, Claude Mythingos Preview, calling the rollout "unprecedented."
• The model, trained to be good at code, developed a side-effect capability in cybersecurity.
• Initial access is limited to ~40 select tech partners, including Apple, Google, Microsoft, Nvidia, and AWS (Project Glasswing), to give them a head start against potential bad actors.
• Anthropic claims the model has autonomously discovered thousands of high-severity vulnerabilities in every major OS and web browser, including some decades-old flaws.
• The company warns the model could be used against corporate networks, healthcare systems, energy infrastructure, and in state-sponsored attacks.
• Initial reports of the model's capabilities hit cybersecurity stocks, but partners in the rollout are now seeing a boost.
• JPMorgan highlighted partners Palo Alto Networks and CrowdStrike as beneficiaries, calling the model a "significant leap forward" and stating cyber names are "essential layers in the defense stack" rather than competitive targets.
• Anthropic describes the effort as "fighting AI with AI" and has already spoken to US government officials about collaboration.
• The announcement comes amid a tense legal battle between Anthropic and the Pentagon over its designation as a supply chain risk.