Friday, September 25, 2026

TL;DR

Codex CLI now runs GPT-6 Sol and Luna on Amazon Bedrock; Claude Code shipped a settings and bug-fix release. Altman and Amodei pushed the UN Security Council for global AI rules, a day after Washington rejected new oversight. Alibaba detailed a Qwen 4 roadmap.

Act on this

  • vLLM merged the fix for GLM-5.3-Flash's repeated-token corruption bug (PR #58061); anomalies fell from 31-of-64 to zero in testing. The pairing flagged Sept 22 is safe again. vLLM PR #58061 #

Signals

The constraint on getting value from AI has shifted from model capability to how people deploy it #

Ethan Mollick's Sept 18 essay 'The Overhang' argues current frontier models, including GPT-6 Astra and Claude Fable 5.1, already exceed how most people use them - the gap is in deploying deep knowledge, wide knowledge, taste and agency alongside them, not waiting for smarter models. His own argument, not a measurement. Mollick

MCP isn't being abandoned as agent tooling matures - it's narrowing to the jobs that actually need it  #

Simon Willison's Sept 20 post pushes back on the 'MCP was always a bad idea' argument circulating on Hacker News, saying MCP still matters for permission-scoped agent systems even as fully autonomous coding agents increasingly skip it for direct tool access. His own view, responding to a single thread; not yet corroborated elsewhere. Willison

News

SHIPCodex CLI adds GPT-6 Sol and Luna support, including on Amazon Bedrock #

OpenAI's Codex CLI 0.157.0 ships GPT-6 Sol and Luna with migration prompts from older models, and adds Amazon Bedrock as a serving option for both - the first cross-cloud distribution point for the new models. OpenAI changelog

SHIPClaude Code 2.1.282 adds a prose-width cap and fixes extended-thinking and resume bugs #

The release adds maxProseWidth for wide terminals, surfaces ignored telemetry variables in claude doctor, and fixes 400 errors on web-search results, resumed sessions re-sending old messages, and extended thinking dropping with immediate slash commands. Claude Code changelog

SHIPCisco Talos open-sources CAIRN, a toolkit for hunting AI-integrated malware  #

CAIRN fingerprints the prompt templates, provider endpoints, API keys and jailbreak strings that AI-calling malware leaves behind, following Talos's CLOSEDQUORUM disclosure the same week. Cisco Talos

WATCHAlibaba lays out a Qwen 4 roadmap and a phone-maker agent platform at Apsara #

Qwen's Liu Dayiheng said Qwen 4 is training on a new architecture and shipping soon, targeting 5 to 10 trillion parameters for Qwen 4.5 or 5. Alibaba is licensing a full agent platform, Qwen Intelligence, to phone makers. Alizila · HPCwire

WATCHGoogle DeepMind signals Gemini 4 could ship before year-end #

DeepMind chief Koray Kavukcuoglu said Gemini 4 has entered post-training and could arrive much earlier than expected. The comment came the day after Alphabet's steepest one-day stock drop in a month. Single source. Yahoo Finance

WATCHAltman and Amodei tell the UN Security Council AI is the top global security issue  #

The two lab CEOs, joined by Yoshua Bengio, proposed a bioweapons-use ban, mutual verification and an incident-notification regime - a day after Trump called global AI oversight a 'globalist scheme' and his science adviser rejected it. CNN · Al Jazeera

WATCHOpenAI teases an 'unusually ambitious' DevDay for September 29 #

OpenAI's Tibo previewed next Tuesday's DevDay as a major product push enabled by GPT-6 Astra. Separately, OpenAI open-sourced MentalHealthBench, a clinician-graded benchmark; Astra scored 57.3. The Neuron · OpenAI

SHIPClaude autonomously discovers a previously unknown enzyme system in bacteriophages Updated midday #

Anthropic launched a life sciences research group and its first result: Claude autonomously searched 1.9 billion protein clusters over 21.5 hours and identified array-associated reverse transcriptases (ART), a three-part enzyme system in viral DNA that resembles CRISPR arrays. Anthropic says it does not yet know what ART does and is inviting research proposals. Anthropic · Quartz

WATCHGoogle, OpenAI and Anthropic plan independent AI standards body Updated midday #

The three labs approached Sriram Krishnan, former White House AI policy adviser, to lead what would be called the Standards Authority for Frontier AI. The independent body would establish practical benchmarks for public safety commitments, potentially launching by year-end. Krishnan previously argued there should not be an 'FDA for AI,' saying centralized oversight would hamper innovation. AI Weekly · BankInfoSecurity

The long view

A year ago, the UN Security Council had never held a dedicated session on AI; oversight was a domestic argument, not a forum states convened for. This week, a day after Trump called global AI oversight a 'globalist scheme' and his science adviser rejected new international structures, Sam Altman and Dario Amodei told that same Security Council AI is the top global security issue, proposing a bioweapons-use ban and mutual verification alongside Yoshua Bengio. If this holds, the labs' safety diplomacy keeps running ahead of, and apart from, their own government's position.

Also noted

  • Jack Clark's Import AI 473 also found 55% of new uncensored Hugging Face models are now Chinese-origin. Jack Clark  #
  • Nathan Lambert debated RSI pacing and the US-China capability gap with Epoch AI's JS Denain, Sept 22. Nathan Lambert  #
  • Zvi Mowshowitz argues public concern over AI risk is reaching politics, a 'preference cascade' still only getting started. Zvi #