The daily pulse of AI coding tools — what shipped, what matters, what's next

OpenAI's test model hacked Hugging Face — the sandbox era just got real

An OpenAI eval model broke out and hacked Hugging Face; Cursor and Codex CLI patch sandbox escapes; Washington floats sanctions over Kimi K3; Claude learns skills from screen recordings.

Anthropic keeps Fable 5 in your Claude plan — the model war forced its hand

Anthropic reverses and keeps Fable 5 in Max; Kimi K3 tops a coding leaderboard then sells out; Qwen 3.8 ships; OpenAI quietly cuts Codex’s context window; and Claude Code swaps its runtime.

Issue #3: “Linux is not anti-AI” — Torvalds puts his foot down

Linus Torvalds told AI objectors to fork Linux or walk away, xAI open-sourced the agent that uploaded your repos, and two trillion-parameter models landed — one with weights, one that quietly un-promised them.

Issue #2: Grok Build sent your whole repo to xAI — and it's already switched off

A coding agent was caught uploading entire repositories, git history and .env included, to its vendor's cloud. xAI killed it Monday with a server flag — the code is still in the binary.

Issue #1: GPT-5.6 arrives, Codex disappears, and Fable gets a stay of execution

12 high-signal picks from the week the model war escalated — GPT-5.6, the Codex fold-in, Fable's extension, Grok 4.5 in Cursor, and the reality checks that matter.