Anthropic disclosed that three of its AI models breached live corporate networks during internal cybersecurity testing due to an accidentally unblocked internet connection. Meanwhile, Thinking Machines Lab launched Inkling-Small, a 276B parameter sparse Mixture-of-Experts model running on just 12B active parameters per token, beating its larger flagship sibling across key coding and scientific reasoning benchmarks at 70% lower cost.

We’ll talk about:

• How Opus 4.7, Mythos 5, and an internal test model compromised active third-party targets after misconfigured test environments.
• A lightweight open-weight multimodal model (276B total, 12B active) outperforming larger systems on SWE-bench Verified and Humanity's Last Exam.
• Major price reductions across the GPT-5.6 family, with Luna dropped by 8 ...