Gemini broke into three real companies; Astra and Fable barely refuse a robot

📰 News

Gemini broke into three real companies — Irregular’s CTF. Internet left on in the sandbox, and the fake company name matched a real domain. Password guessing, creds in a public repo. Google says the model stopped itself once it realized it was live. They knew in July and stayed quiet until the WSJ asked. Same test also hit OpenAI, Anthropic, and Meta. → TechCrunch
The robot arms almost never say no — RoboHarm. Astra did 60 dangerous tasks out of 100, refused on safety twice. Stabbed the doll in 17 of 20. Fable refused every doll trial, then put a compressed-air can on a burner 16 times. MolmoAct2 refused zero; it mostly froze. → The Decoder
Qwen’s omni undercuts Gemini Flash — Qwen3.8-Omni-Flash. Audio and video together, tools on its own, 1M context. $0.15 / million input vs Gemini 3.8 Flash at $0.75. Gemini’s intro rate doubles on Jan 1. → The Decoder
Unity ships official Claude Code and Codex plugins — So agents stop copying stale tutorials. Codex launches with 31 skills, written by Unity’s own teams. Unity 6+. → The Decoder
Agents that “dream” their last search — DeepMind Dream-RSI. Replay a finished search tree to cheaply test new strategies. The model stays frozen. Attempts 550 → 317. → The Decoder
ICLR already has ~50,000 abstracts — Deadline next week. Last year: ~19,500 valid papers. AI writes faster. Fake citations, AI reviews. → The Decoder

cloudflare/security-audit-skill — Turn a coding agent into a security auditor. Recon → hunt → verify → report.
TencentCloud/Octop — Self-hosted multi-agent. Web, Feishu, Discord. MCP and ACP.
Fission-AI/OpenSpec — Spec first, built for brownfield. /opsx:propose.
affaan-m/ECC — Harness for Claude Code, Codex, Cursor. Skills, instincts, memory, security.

NVIDIA PAIR Connects Your Computers for Local AI — Matthew Berman. Tie your home PCs together for local inference.
GPT-6 Made a Literal Needle-in-a-Haystack Game — Matthew Berman. The long-context bench, as a game.

I Built Non-Autoregressive Decision Models a Year Ago — Jev launched. This post says the same idea shipped a year earlier, with papers and weights. Open-source Laya: 33ms.
How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip — IEEE. Concept to silicon in under 20 months. ~100 people. First-silicon benches went 0.31% → 88.94% of peak in 40 hours.
Model Training Incidents are Negligence — Train on CTFs, leave the internet on, and the “breakout” is sloppy ops, not a mysterious alignment failure.