June 5 is a split-screen day for AI infrastructure: the model race keeps compounding (Microsoft ships seven new MAI models at Build, GPT-5.5/5.4 and Codex land on AWS Bedrock, Google ships an encoder-free Gemma 4 12B) while the cost and safety reckoning gets louder. Uber is now capping employee AI-tool spend at $1,500/month, and Anthropic published a rare deep-dive on how it technically contains Claude even as it pushes the model deeper into critical infrastructure across 15+ countries. On the security side, agentic tools are cutting both ways: OpenAI's Codex agent found a real HTTP/2 DoS bug in the wild, while University of Toronto researchers demoed an AI worm and GitHub.dev disclosed a one-click OAuth token theft flaw. For engineers, the throughline is that agentic AI is now expensive, embedded in high-stakes systems, and itself a new attack surface - all at once.

• Uber caps AI coding tool spend at $1,500/employee/month after costs spiked
• Microsoft launches seven MAI models at Build, including agentic MAI-Code-1-Flash
• GPT-5.5, GPT-5.4 and Codex now available on AWS Bedrock
• Anthropic details Claude's containment architecture while scaling it into critical infrastructure in 15+ countries
• Codex agent discovered a hidden HTTP/2 denial-of-service bug