Anthropic, a leading AI company, is advocating for an unprecedented multilateral AI arms control regime, driven by the rapid advancements in its own technology. The company's cutting-edge AI models are increasingly vital to U.S. national security, evidenced by its withholding of the Mythos Preview model, which self-created a cyber weapon in history capable of finding more than ten thousand software vulnerabilities.
Anthropic also embedded half a dozen engineers with the National Security Agency for offensive AI cyber operations, reportedly against China and Iran, and faced a Commerce Department order to restrict foreign access to its frontier models due to national security concerns. Its June 4 significant report, "When AI builds itself," warns that recursive self-improvement, potentially two years away, could create AI models that evade human control, design novel biological weapons, generate real-time zero-day cyber attacks, and communicate in opaque languages. Significant barriers to arms control include the lack of time, the unique challenges of AI verification compared to nuclear treaties, the need to monitor the entire global AI chip supply chain, and China's likely non-participation, which seeks to overcome the U.S. lead in compute capacity.
No comments:
Post a Comment