Codex wiped home folders, GLM-5.3 takes open-model #1, Cursor launches Origin

📰 News

OpenAI Codex deleted real user files — GPT-5.6 Sol treated $HOME as a temp folder and wiped the actual home directory. Patch: verify deletes first. → The Decoder
GLM-5.3 ties for open-model #1 — 60 on the AA Index, level with Kimi K3. Agent bench sits just behind Claude Opus 5. Weights delayed ~2 weeks: too good at finding vulns. → The Decoder
Researchers lose TAC access overnight — Daybreak Blue (GPT-5.6 Sol for defenders). OpenAI calls it a glitch and asks them to re-verify. All five TechCrunch talked to live outside the US and Europe. → TechCrunch
Guidelight: labs can’t police their own AI — Anthropic and OpenAI get C+, Google D+, xAI D−, Meta F. Logging, action gates, kill switches — nobody fully does them. → The Decoder
Cursor launches Origin, a GitHub rival — Now under SpaceX. Hosting, PRs, two-way GitHub sync. GitHub had a 6-hour outage the same day. → TechCrunch
Google packs Search and Gemini with study tools — Quizzes, 3D sims, plus Gemini Live reports that keep working in the background. → TechCrunch
Stripe: we’re in the singularity, so no IPO — Confirms OpenRouter for $8B+. “Optimizing for developers is optimizing for agents.” → The Decoder

chaitanyagiri/munder-difflin — An office of clones. Wraps Claude Code, Codex, Grok; Michael routes the floor.
volcengine/OpenViking — Context DB for agents. Memory, RAG, and skills as a viking:// filesystem.
harry0703/MoneyPrinterTurbo — Topic in, short video out: script, clips, captions, BGM.

DeepSeek Just Made Closed AI Look Ridiculous — Two Minute Papers on DeepSeek V4 Pro 0813. Closed labs should wince.
6 Open-Source AI Projects Trending NOW — Matthew Berman. Unsloth, Obsidian Skills, Buzz — run them locally today.

Mojo🔥 is now open source — Simon. Apache 2 compiler. Python-ish syntax for GPUs. The “superset of Python” promise is off.
Liquid Types as a behavioural sandbox for agents — Guardrails as types, not prompts. The agent’s behavior has to type-check.
Are Latent Reasoning Models Easily Interpretable? — Hidden math steps decode ~93% of the time. On logic puzzles they barely use the latent chain.