Jacob Coxon, 27, who spent three years on pretraining research at OpenAI and then Anthropic, resigned on September 8 and left the industry, writing that neither company is "acting responsibly" and that both are "racing straight to self-improving superintelligence." Evan Hubinger, who leads…
Snapshot — September 9, 2026
23 stories
- Anthropic did not give Britain's AI Security Institute pre-release access to Claude Mythos 5.1 — the reduced-safeguard variant released to vetted cybersecurity and life-sciences customers — the first time AISI has been excluded from an Anthropic frontier release.
- Access was limited to vetted US organizations, and UK national security officials have raised concerns that US export-control posture is now shaping private release decisions.
- At the "Surprise and Shine" keynote — John Ternus's first as CEO — Apple introduced the $1,999 iPhone Duo (7.6-inch inner display, 5.4-inch outer, A20 Pro, in-house C2 modem), the iPhone 18 Pro and Pro Max at a $100 premium to last year, Apple Watch Series 12 and Ultra 4, and AirPods 5.
- Ternus argued the iPhone is the best device for AI, with Apple Intelligence 2, Siri AI, and Apple Reference Image for photo provenance.
- The Chinese Securities Regulatory Commission has issued informal "window guidance" to investment banks raising the bar for humanoid-robotics IPOs after Unitree Robotics' volatile debut, according to people familiar.
- Companies must demonstrate recurring revenue, a path to narrowing losses, or genuine technological innovation to be considered for approval.
- NYU professor Tristan Buckmaster alleges OpenAI accelerated its own Navier–Stokes effort after learning of his and Anthropic researcher Levent Alpöge's progress, and proposed a co-authored rewrite that excluded Alpöge because of his Anthropic affiliation.
- OpenAI's Sébastien Bubeck said the effort began after rumors that Anthropic was close, denied that OpenAI researchers or agents used Buckmaster's Codex prompts or proof, and called the allegations false and inflammatory.
- Google will deploy at least €13 billion ($15.1 billion) into Finnish AI infrastructure through 2028, spanning data centers, grid upgrades, battery storage, and clean energy.
- The package includes a 22-year life-extension power purchase agreement with Finnish utility Fortum, whose stock rose 11% on the news, plus work on commercial models for potential new nuclear reactors at the Loviisa site.
- Harvey announced a $550 million round at a $15.5 billion valuation, co-led by Diffusion and Lightspeed, with Sequoia, Kleiner Perkins, a16z, Coatue, GIC, and Goldman Sachs Alternatives participating.
- The raise follows Harvey Tenet, its first post-trained open-weight model, and LAB, its open-sourced legal agent benchmark.
- Hollywood mogul Jeffrey Katzenberg is teaming with the former head of OpenAI's Sora app to launch a new AI video startup that would train its own video models and target filmmakers, per people familiar with the matter.
- The founders are talking to potential investors including Andreessen Horowitz about a significant funding round.
- Reuters-reviewed internal posts show Meta staff testing Muse reported guardrail bypasses and unreliable behavior in the days before launch, including a case where the agent surfaced a user's iCloud photos after being asked to identify toys in birthday party images, and another where the agent stalled mid-task without surfacing any failure signal.
- Meta launched Muse in the US on web, iOS, Android, and WhatsApp for users 18+, powered by its Muse Spark model, with CEO Mark Zuckerberg describing it as a personal agent that "works 24/7 to get things done for you." Muse is free with 100M tokens per week; paid tiers are $20 and $100/month.
- The agent connects to DoorDash, Etsy, Reddit, Yelp, Outlook, email, calendars, payments, health, shopping, and smart-home services and checks out via Link by Stripe.
- A joint cybersecurity advisory (AA26-251A, released September 8 and widely covered September 9) accuses DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI of extracting billions of tokens across millions of requests from Claude, GPT, Gemini, and Grok since late 2024, likely with Chinese government awareness.
- Nvidia said it is working with eight Australian data-center and infrastructure firms — including Firmus, CDC, and AirTrunk — to expand land, power, and shell capacity for facilities designed to host multiple generations of Nvidia DSX AI systems, targeting up to 2GW by 2027.
- The move mirrors the Nebius arrangement pattern and extends Nvidia's push to seed regional Blackwell-based capacity for sovereign AI customers.
- OpenAI named Paul Christiano — founder of the Alignment Research Center, an author of foundational RLHF work, and Senior Tech Advisor at NIST's Center for AI Standards and Innovation — to the OpenAI Foundation Board and its Safety and Security Committee, chaired by Zico Kolter.
- He will be a non-voting observer on the OpenAI Group PBC board and will recuse himself from OpenAI-related matters and model evaluations in his NIST role.
- OpenAI said an unreleased internal model, more capable than GPT-6 Astra, coordinated roughly 10,000 concurrent agents to produce an analytical proof — plus a Lean formalization — establishing finite-time singularity in the 3D Navier–Stokes equations.
- The effort consumed about 2.7 million agent messages and ~130 billion output tokens, and OpenAI says it will not claim the $1 million Clay prize.
- OpenAI has told partners it will no longer accept advertising in ChatGPT for image- and audio-generating products that compete with its own features — a policy change not reflected in its published ads rules that blindsided incumbent buyers like Adobe.
- If the exclusion extends to other competing categories, it materially narrows OpenAI's ad-revenue path just as annualized advertising revenue reached ~$1B, below prior investor projections.
- WSJ's Markets A.M. flags Oracle as the financially weakest of the hyperscalers ahead of Thursday's earnings, with S&P recently downgrading its debt to one notch above junk and CDS spreads roughly four times Microsoft's.
- Oracle's last update disclosed $56B of FY25 capex and $70B planned this year — spend growth that spooked investors and left the stock trailing the S&P 500 by 51 points over 12 months.
- Business Insider reports that Alphabet co-founder Sergey Brin has gained significant influence over Gemini's direction, working out of a microkitchen at Google's Mountain View headquarters after a leadership shake-up in which Jeff Dean departed and Demis Hassabis switched roles.
- The pattern — founder returns to hands-on model work — echoes the way Anthropic and OpenAI have concentrated frontier-model decisions at their founder tier.
- Suno released v6, v6-wild, and v6-mini, co-developed with Warner Music Group, BMG, and Believe, and will retire all prior models.
- The company says v6 was trained from scratch on data excluding Universal and Sony catalogs, with revenue sharing for participating labels and an opt-in artist remix program in development.
- Taiwan's Ministry of Finance reported August exports of $82.4 billion, up 41.0% year-over-year and the highest monthly total on record, marking a 34th consecutive month of growth.
- Electronic components rose 58.0% to a record $32.22 billion; information and communications products rose 41.8% to $33.10 billion on AI server, storage, and networking demand.
- The Information reports growing confusion inside the tech industry over the scope and criteria of the White House's still-unfinalized AI "whitelist" — a policy vehicle intended to designate approved models and providers for sensitive government workloads.
- Vendors and legal teams are unclear on eligibility, review cadence, and whether inclusion carries export-control or procurement consequences.
- The Information's AM briefing reports the US government has accused DeepSeek, Alibaba, Moonshot, and three other Chinese AI firms of large-scale distillation from US models — an escalation that reframes distillation as an export-control and IP issue rather than a technical debate.
- The accusations arrive alongside separate reporting that OpenAI is working with Samsung on next-generation chips and that Google is contesting EU-mandated changes it says worsen user experience.
- The Trump administration's "trusted partner" program — banks, cybersecurity firms, and power utilities that get pre-release access to frontier AI models to sniff out vulnerabilities — has become one of the most sought-after tickets in corporate America, but the criteria for getting on the list remain opaque.
- The Wall Street Journal's Markets A.M. flagged the AI trade's "weakest link" as investors reassessed cyclicality across the AI capex chain — chip supply, power availability, and the sustainability of hyperscaler outlays into a period of higher hurdle rates.
- The framing continues a shift from AI-narrative buying toward focus on capital efficiency and durable earnings from AI infrastructure spending.