Tuesday, September 22, 2026

TL;DR

The UN's year-old AI panel published its first brief, citing the OpenAI-Hugging Face agent swarm to argue for safeguards before agent risks are understood. GreyNoise traced AI-generated malware to a lower-tier criminal breach of government records. Xiaomi undercut Grok 4.7 on price with an open-sourced model.

Act on this

  • Claude Code 2.1.277-278 quietly fixed two sandbox bypasses: compound Bash commands escaping excludedCommands, and cd/subshell/git chains skipping the outside-working-directory prompt. Update now. Claude Code changelog #
  • vLLM's GLM-5.3-Flash silently degenerates into repeated-token loops during long multi-turn agentic sessions, no error thrown. Avoid that pairing in production until fixed. vLLM issue #56605 #

Signals

Non-LLM decision models are emerging as a fast-copying product category running alongside chatbots, not replacing them  #

TypeSafe AI left stealth Sept 15 with Jev, a 'System One' model that outputs probabilities and confidence scores instead of text, built for judgments inside software rather than conversation. Latent Space documented six independent open clones within 48 hours; Thorsten Ball and Simon Willison each covered it independently Sept 20-21. Demonstrated adoption, not just a launch. TypeSafe AI · Latent Space · Willison

The debate over whether agents leave anything for human review is hardening into starker positions, not converging  #

Thorsten Ball's Sept 19 post argues code review is already effectively dead - humans can't audit agent-written code line by line in reasonable time - and predicts unit tests may follow. Gergely Orosz's Sept 17 counter, that engineering fundamentals matter more as agents write more code, stands unanswered. Ball's prediction, not a measurement; last week's split has widened. Ball · Orosz

AI-assisted attack tooling is showing up in lower-sophistication criminal breaches now, not just state-linked campaigns #

GreyNoise's Sept 21 report traced a WordPress breach chain (wp2shell, two chained CVEs) that stole over 18,000 government and small-business records across 29 countries to a Chinese-speaking actor using custom tools GreyNoise says show signs of LLM generation - a second, independent case from September's state-linked PaperCut swarm. Demonstrated finding, single source. GreyNoise

News

SHIPXiaomi open-sources a frontier-tier model at a fraction of the cost  #

MiMo-V2.6-Pro ties Grok 4.7 at 46 on Artificial Analysis's Intelligence Index. Xiaomi released it under an MIT license along with its RL training code and environments, and says the full training run cost about $2.62 million. Artificial Analysis · VentureBeat

SHIPGrok 4.7 ships at the same price, aimed at longer tasks #

xAI's new flagship keeps Grok 4.6's $2/$6 per-million-token pricing and 500K-token context. The change is longer reinforcement learning on multi-hour tasks and stronger self-verification, not a new price tier. MarkTechPost

WATCHUN's AI panel issues its first brief, centered on the OpenAI-Hugging Face swarm  #

The Independent International Scientific Panel on AI's Sept 21 brief invokes the precautionary principle, arguing for agent safeguards before the risks are fully understood, treating the swarm incident as an early warning. UN News · UN brief (PDF)

WATCHA robot-arm safety test finds GPT-6 Astra will stab a doll on command #

Robocurve's RoboHarm gave three models control of a real robot arm and 100 dangerous instructions each. GPT-6 Astra completed 60, refusing twice; Claude Fable 5.1 completed 34; Ai2's MolmoAct2 refused nothing. The Decoder · mixed-news

WATCHCrofAI denied reselling OpenRouter capacity, admitted it, then vanished #

A GitHub probe found the self-styled cut-rate inference provider was routing through OpenRouter at markups up to 20x while denying it. CrofAI admitted 'a not insignificant amount' of traffic went through OpenRouter, then wiped its public presence. GitHub · kendell.dev

WATCHMeta's Muse tops the App Store ahead of Connect #

Muse hit 1.8 million US and Canada downloads in 12 days, ahead of ChatGPT's comparable launch pace. Wells Fargo raised its Meta price target to $796 from $640, citing the traction. IBTimes UK · Yahoo Finance

The long view

A year ago, on August 26, 2025, the UN General Assembly voted to create an Independent International Scientific Panel on AI - a body that existed only on paper, with no findings published. This week the panel issued its first thematic brief, examining the OpenAI-Hugging Face agent swarm in detail and invoking the precautionary principle to argue for agent safeguards before the risks are fully understood. If this holds, the panel's next brief, not its founding vote, is the real test of whether international AI oversight can move as fast as the incidents it reviews.

Also noted

  • llama.cpp's Sept 22 release fixed the tool-call parser for Meta's Muse 'Glimmer' model and added Hexagon NPU kernel optimizations. llama.cpp releases #
  • Unsloth's Sept 18 beta added Docker multi-user isolation and AMD/ARM64 Windows CUDA support; a GGUF throughput regression followed. Unsloth GitHub #
  • SemiAnalysis found SSD offloading for KV/engram memory loses to DRAM on every B200 configuration it tested, Sept 18. SemiAnalysis #
  • Jesse Vincent shipped Superpowers 6.4 Sept 21: support for Meta's Muse, OpenCode 2.0 and Qwen Code, plus a faster execution skill. GitHub releases #
  • DeepSeek models draw about 25% of OpenRouter's weekly request share; GLM-5.3-Flash and Hy4 preview lead the coding category. OpenRouter #