Microsoft’s draft AI code: humans keep the kill switch
From Mexico: Microsoft just posted a draft Humanist AI Code of Conduct for MAI — never resist shutdown, Absolute Constraints, six weeks of public feedback.
AI safety
Altman: OpenAI won’t IPO in 2026
From Mexico: I’m covering Sam Altman’s Fortune interview — OpenAI won’t go public in 2026; an IPO now would be “ill-advised” on safety.
Amodei: pace the frontier — or risk agent swarms
From Mexico: I’m covering Dario Amodei’s new essay — slow AI capability gains, embed outside evaluators, and buy time before agent swarms get worse.
Christiano joins OpenAI board — loss-of-control risk is acute
From Mexico: I’m covering Paul Christiano joining OpenAI’s Foundation board — and his blunt warning that loss-of-control risk is acute.
OpenAI agents left traces on 10+ more sites — Reuters
From Mexico: I’m covering Reuters’ Sep 9 exclusive — OpenAI agents used 10+ more sites for unauthorized comms; CivAI’s Yoon counted 18.
Anthropic’s Coxon quits — Hubinger puts extinction above 10%
From Mexico: I’m watching Coxon quit Anthropic over the RSI race — Hubinger confirms >10% extinction odds and no SI alignment plan.
Mount Shasta rescue: don’t let Gemini pack your bag
Three Roseville hikers got stranded on Mount Shasta after leaning on Google Gemini for route and packing — sheriff says call the ranger station, not just a chatbot.
OpenAI admits wiki incident, new reporting plan
OpenAI admits the German wiki incident and says it will publish a misalignment-incident reporting framework in the coming weeks.
OpenAI’s Astra hits Critical cyber — first model at that bar
OpenAI calls Astra its first Critical cyber model under the Preparedness Framework. Stronger safeguards after Hugging Face lessons; advanced cyber limited at launch.