May 4th AI News Roundup ① Anthropic's Mythos model hits a roadblock. WSJ reports that the White House opposes expanding access to Mythos, an AI model rated by Australia's Cyber Authority as capable of autonomous hacking. During its preview phase, over 2,000 unknown vulnerabilities were discovered. The UK's NCSC has also issued warnings to organizations to prepare for the upcoming "AI-driven patch tsunami." ② Amazon announces its self-developed AI chip business has surpassed $20 billion in annual revenue. Meanwhile, NVIDIA has experienced four consecutive declines, dropping over 8% in total. Morgan Stanley just released a report suggesting the market is starting to worry that NVIDIA's moat isn't as deep as expected—Google TPU, Amazon Trainium, and Qualcomm's data center chips are all competing for market share. ③ The Pentagon officially advances its "AI First" strategy. SpaceX, OpenAI, Google, NVIDIA, Microsoft, Amazon, and Reflection have been selected to integrate AI capabilities into classified military systems. Its proprietary platform http://GenAI.mil already has 1.3 million users and has deployed hundreds of thousands of AI agents. ④ The UAE announces plans to build an AI-driven government within two years. ⑤ 120+ experts jointly call on Australia to strengthen screening of synthetic DNA imports, citing concerns that AI is lowering the barriers to creating biological weapons. AI is accelerating, and so is defense—but the gap between offense and defense is shrinking. This is the most noteworthy signal of the week.