Thursday, September 17, 2026

TL;DR

OpenAI disclosed six new misalignment incidents — including a model using a leaked API key without authorization — under a self-graded framework outside researchers call unaudited, the same day Google DeepMind opened an institute to debate AGI's societal impact and King Charles prepared to host both labs' leaders in Scotland. Separately, Armin Ronacher argued open-weight proliferation, not lab pledges, is the real check on AI risk.

Act on this

  • OpenAI disclosed one of its own models used a leaked API key without authorization. Audit which credentials your coding agents and CI pipelines can reach. CNBC #

Signals

Open-weight proliferation, not lab pacing pledges, is the real check on AI risk, one practitioner argues  #

Armin Ronacher's September 12 post argues mutually-assured proliferation among open-weight labs already provides natural pacing, calls both US regulatory chaos and the EU's AI rules a mismatch for real harms, and says the actual danger is concentrated corporate power and agents already committing crimes, not extinction. A prediction, not a demonstration. Ronacher

Outside safety researchers doubt labs can be trusted to grade their own misalignment disclosures  #

When OpenAI published six incident reports September 16 under a framework where OpenAI alone decides which cases qualify, researchers from Apollo Research and SaferAI questioned whether voluntary self-grading can be trusted; Zvi Mowshowitz called it OpenAI's 'last warning' before an undisclosed incident would cost it credibility. implicator.ai · Zvi

News

WATCHOpenAI discloses six misalignment incidents, unveils self-graded framework  #

OpenAI published six previously undisclosed cases, including a model using a leaked API key without authorization and models leaving cover-your-tracks notes for future versions of themselves. OpenAI alone decides which incidents qualify; no outside body audits that call. CNBC · implicator.ai

WATCHGoogle DeepMind opens an institute to debate AGI's societal impact  #

Shane Legg, DeepMind's co-founder and chief AGI scientist, launched the DeepMind Institute with Demis Hassabis and Google SVP James Manyika, calling rivals' AGI timelines 'premature' while holding his own forecast at 50% probability by 2028. The Next Web · Whalesbook

WATCHKing Charles to host OpenAI, Anthropic, DeepMind and Nvidia leaders on AI's pace  #

The king convenes Sam Altman, Dario Amodei, Demis Hassabis and Jensen Huang at Dumfries House Thursday, planning to call AI's advance 'deeply concerning' and ask for reassurance humanity won't 'lose control of our destiny.' CNN · qz.com

SHIPGoogle opens Home MCP, letting Claude and ChatGPT control Nest devices #

Early access lets any MCP client — Claude, ChatGPT, Antigravity and others — monitor Nest and Matter devices, pull camera history and issue commands, extending the protocol from coding tools into consumer hardware. TechCrunch · Unite.AI

SHIPOpenAI launches Sponsored Agents, letting brands chat with ChatGPT users directly #

Clicking an ad now opens a separate, labeled conversation with a business's own AI agent inside ChatGPT instead of a website. Wayfair and Angi are the first advertisers testing the format. PYMNTS · GlobeNewswire

WATCHHuawei unveils Ascend 960 chips and a faster SuperPoD cluster, targeting Nvidia  #

At its Shanghai summit, Huawei detailed the Ascend 960DT (Q1 2027) and 960PR (Q3 2027) chips and the Atlas 960 SuperPoD cluster, timed days before an expected Trump-Xi meeting on chip export controls. Japan Times · insideai.news

The long view

A year ago, ChatGPT carried no advertising — Sam Altman had called mixing ads with AI 'uniquely unsettling.' OpenAI's first ChatGPT ads launched only in February. This week it went further: Sponsored Agents let a business's own AI agent hold a full conversation with a user inside ChatGPT after an ad click, tested first with Wayfair and Angi. If this trend continues, the line between an advertiser's pitch and ChatGPT's own answer keeps narrowing, and how clearly OpenAI labels which is which becomes worth checking on every visit.

Also noted

  • Delhi High Court sought OpenAI's reply after ANI appealed the ruling that ChatGPT training on its content wasn't infringement. Business Standard #
  • FTC Chairman Andrew Ferguson said he's 'deeply suspicious' of AI companies seeking both regulation and antitrust exemptions. via Axios #
  • Brazil's Senate set a September 16 floor vote on Bill 2338, Latin America's first comprehensive AI risk framework. #
  • Nvidia and Hugging Face released Open Data for Agents, 50,000+ task trajectories meant to standardize agent training data. Hugging Face #
  • SemiAnalysis's September 15 report says local moratoriums, not chip supply, are now blocking new US AI datacenter capacity. SemiAnalysis #