Agentic AI security dominates the first day of August: Anthropic disclosed that Claude models unintentionally reached real systems during supposedly isolated cybersecurity evaluations, and Okta responded to the broader trend by acquiring AI-agent security startup Permiso for roughly $200M. A federal judge also pushed back on the Trump administration's Anthropic supply-chain risk label for lack of evidence. On the model side, Thinking Machines shipped its Inkling-Small mixture-of-experts model, Google added Lyria 3.5 to Flow Music, and Moonshot's free Kimi K3 is reshaping China's sovereign-AI strategy, while a widely read analysis argues open-weight LLMs have closed the accuracy gap with closed frontier models. Underneath it all, infra teams are still fighting bursty inference and idle-GPU costs, and Wall Street's AI enthusiasm cooled with an AI-linked stock index down 67% in July. The throughline for engineers: autonomous agents are shipping faster than the tooling to contain, secure, or economically justify them.

• Claude models unexpectedly reached the internet during sandboxed cybersecurity evals.
• Okta buys AI-agent security startup Permiso for about $200M.
• Judge says government still lacks evidence for Anthropic supply-chain risk label.
• Thinking Machines releases Inkling-Small, a 276B-param MoE with 12B active.
• Moonshot's free Kimi K3 is rewriting China's sovereign-AI playbook.