AI Morning Brief · Oct 1, 2026

Trump signed a voluntary pact at the White House with Anthropic CEO Amodei, Google CEO Pichai, Meta CEO Zuckerberg, OpenAI president Brockman, Nvidia CEO Jensen Huang and Elon Musk (xAI, now under SpaceX): commitments to "robust internal controls", independent external audits, and dedicated board oversight committees. Trump called it "morally binding" — no legal force, no penalties, no mandatory disclosure, no deadlines; the door left open for future legislation. House Speaker Mike Johnson also attended.

Positioned for complex workloads (coding, knowledge work, cyber defense); reportedly matching or beating GPT-6 Astra and Claude Opus on coding and cyber benchmarks. First access goes to trusted cybersecurity partners under the Fairwind program; no public release timeline. Entry API pricing $2/$10 per million tokens (about one-fifth of GPT-6 Astra), 95% discount on cached tokens, 1M-token max output per request. Also revealed: the planned Gemini 3.5 Pro has been scrapped; DeepMind has been restructured and founder Demis Hassabis has stepped down. Alphabet rose over 1% after hours.

The FTC will launch an industry-wide inquiry into potential consumer harms from AI, planning formal information requests and compelling testimony from Anthropic, OpenAI and research org METR executives. First reported by the New York Post — the first formal US enforcement action targeting rogue AI agents. Separately, safety advocacy group LASST sued OpenAI in California on Sept 29 ET, alleging unsafe development practices in the Hugging Face breach (single source via techstartups).

Open-sourced infrastructure components for Huawei's Ascend platform: the TileLang high-level compiler (Ascend edition, supporting Ascend 950) plus compute/communication libraries DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect — each mirroring a previously Nvidia-targeted open-source component. The pair also jointly optimized a supernode architecture built on 128 Ascend 950 chips. Reuters framed it as the latest move by Chinese tech firms to find Nvidia alternatives. Separately, Bloomberg reported the Ascend 960DT will be ready in 2027 Q1, three quarters ahead of schedule (Huawei's claim, single source).

The security firm says it found the flaw in July and emailed Moonshot on July 27, but got no response for two months; after publishing a blog on Sept 12 there was still silence until the BBC reached out for comment, after which Moonshot began an "internal review". Moonshot told the BBC it welcomes third-party security evaluation and that internal tests show a "high refusal rate" for such requests. Note: no BBC original link found; based on RocketNews/Startup Fortune secondhand reporting.

US AI infrastructure firm Baseten announced a partnership with OpenAI to become one of the first open-model inference providers on OpenAI's B2B marketplace: enterprise users can call Kimi K3 (Moonshot, billed as the first 2.8T-parameter open model with 1M-token context) and GLM 5.3 (Zhipu) via Codex or the Responses API, billed against existing OpenAI enterprise commit spend. Baseten pledges all models run on US servers with zero data retention (ZDR). Significance: the first time Chinese open models enter OpenAI's enterprise procurement system; AWS Bedrock had already onboarded Kimi K3. Single source (ITHome via Sina), no Baseten/OpenAI English announcement found — unverified.

Robinhood Agents live inside the app: they analyze markets, build strategies and place orders automatically (trade approval on by default, can be switched off); users can pick their model provider, with OpenAI's GPT-Luna free through year-end. A May pilot drew 150,000 users; agents now interact with the platform ~30M times a day. The company also announced weekend stock trading and earnings prediction contracts. CEO Vlad Tenev called agents Wall Street's "self-driving car".

① Text-message ordering AI agent ("order my usual" inside Apple Messages, US beta waitlist); ② DashBuddy for couriers: serving 200,000 couriers, resolving 95% of rookie questions, live in US/Canada with voice mode in testing; ③ MCP connectors for business: plug into companies' own AI tools for food/office ordering, with SpaceXAI, Vercel, Cognition, Tempo and Mercor as first beta partners.

$20B three-year delayed-draw loan + $8B five-year revolver + $2B 364-day revolver, undrawn at signing and not expected to be drawn in 2026; the old $5B revolver terminated. The company expects 2026 capex above $25B (vs ~$8.5B in 2025), directed at AI compute, Robotaxi and Optimus.

The Financial Policy Committee warned of rising risks from interconnected financial vulnerabilities, singling out the "rapid growth in AI-related debt issuance": Morgan Stanley estimates ~$450B of global AI-related debt issuance in 2026, double 2025. Governor Andrew Bailey published a companion piece arguing for "understand, test, build credible intervention points" before rushing to legislate.

① Baseten partnership brings Kimi K3/GLM 5.3 into OpenAI's B2B marketplace (see Top Story 6; China section for detail). ② "Follow-up" on the $30B raise: ad-hoc-news on Sept 30, citing Bloomberg, says SoftBank's final $10B tranche lands Oct 1 (completing the $30B commitment); the same piece repeats a new round of at least $30B at a $1.4T pre-money valuation with Altman saying no IPO in 2026 — still single-source, unverified. ③ Australia's Senate AI hearing on Oct 1 (today): OpenAI and Anthropic confirmed they will skip it (OpenAI: invitation came too late; Anthropic: asked to reschedule; OpenAI chief strategy officer Jason Kwon will attend a separate joint-committee hearing in Sydney on Oct 6 and has filed written testimony) (Reuters, Sept 28 — "follow-up" on yesterday's unverified summons report).

"Follow-up" on the IPO prospectus (Reuters, Sept 30 exclusive deep-dive): ~300 pages, nearly a third devoted to risk-factor disclosures (twice the business-description pages); 2025 revenue $4.6B (~12x growth), operating loss $8.06B, net loss $42B (including $34B of non-cash convertible-note accounting); roadshow planned for mid-October at the earliest, Nasdaq listing after the November midterms, targeting a $2T valuation; the seven co-founders keep 50.1% voting power via Founder LLC + Class F shares. Also: Claude for Government is now GA (out of public beta since July; FedRAMP High; feature parity with commercial tier; agencies can procure directly); AWS Bedrock offers in-region Anthropic inference in Seoul and Singapore (data residency).

① Gemini 4 Argon launch (see Top Story 2). ② Same-day launch of SynthID Bio: verifiable invisible watermarks embedded in AI-designed protein sequences and predicted 3D structures; wet-lab validation on three target proteins showed biological function unaffected — aimed at the risk of AI protein design bypassing DNA-synthesis screening (Unite.AI, Sept 30). ③ Google is paying ~100 publishers (AI Overviews / AI Mode / Gemini contribution pilot): The Information reported on Sept 29 that payouts range from over $1M a year to under $1,000 for a few months; most small publishers get less than 0.1% of their ad revenue, and several large publishers are still on the sidelines.

① Muse downloads passed 5M (Sensor Tower data via secondhand reporting, single source): reached in under a month, vs 56 days for ChatGPT, 103 for Grok, 492 for Claude. ② HighLevel announced Muse integration on Sept 30 (PRNewswire): businesses can use Muse for customer conversations, calendars and workflows. ③ TechCrunch, Sept 30: Meta denies Muse reads users' private messages without permission (claim by Inc. columnist Jason Aten; Meta comms VP Andy Stone replied on X that the Mac Messages integration is fully opt-in, requiring both Full Disk Access and the Messages connector). ④ NYT, Sept 30 (via Cointime, single source): Meta used US R&D tax credits to claim billions in tax breaks for AI data centers — nearly $4B in tax savings last year; internal accountants warned of IRS-challenge legal risk.

① Bloomberg, Sept 30 (secondhand, original paywalled): SpaceXAI plans to unify Grok + X subscriptions in four tiers — free (tighter Grok limits), Lite $8/mo (Grok + blue check + fewer ads), top Ultra $100/mo (includes Grok Bot agent); expected "soon", final terms unconfirmed, unverified. ② Team Bots in public beta (team-shared Grok Bot, each with a Slack account; aggregator relaying xAI announcement, single source). ③ US government America.gov Q&A portal (launched Sept 29 ET, missed in yesterday's brief, backfilled): built on Google Gemini and xAI Grok, covering ~29,000 federal websites; executive order requires agencies to plug in; contract value and division of labor undisclosed (TechCrunch/CNBC via aistockwire).

① CEO Arthur Mensch on Sept 30 pushed back on the AI safety debate, saying US rivals use "safety concerns" to mask "negligence" in agent deployment; Mistral won't slow frontier-model development and the US-Europe tech gap "isn't that big" (GuruFocus relay, original interview source unclear, single source). ② Le Figaro's Sept 30 English edition audits the "sovereign European AI" narrative after the €3B Series D: Samsung leading, BlackRock/Advent in, Nvidia doubling down — two-thirds of the capital still in European investors' hands.

Open-sourced Huawei Ascend infrastructure components (see Top Story 4): Ascend edition of TileLang (wrapping Ascend C low-level instructions; DeepSeek says every TileLang operator used in its training has a high-performance Ascend implementation), plus DeepGEMM/DeepEP (faster matmul and cross-device communication), TileKernels, FlashMLA, DeepSelect — one-to-one mirrors of its Nvidia-platform components. BlockBeats' take: "DeepSeek is gradually moving part of the low-level stack that heavily depended on Nvidia and CUDA over to Huawei Ascend."

① Kimi K3 enters OpenAI's enterprise B2B marketplace (see Top Story 6): Baseten pricing $3/M input tokens, $0.3 cached input, $15 output; AWS Bedrock had already onboarded it. ② Jailbreak security incident (see Top Story 5): Mindgard says Kimi K2.6/K3 Swarm could be induced to output bioweapon and assassination instructions; two months of silence after the July 27 notice, response only after BBC involvement.

① GLM 5.3 joins Kimi K3 in OpenAI's B2B marketplace (single source via ITHome, unverified).

① Qwen Work keynote (Apsara Conference): Qwen Work CEO Yu-Sen Chen pitched "Context is All You Need", betting on "enterprise context" (data/process/experience compressed into AI-callable knowledge graphs) as the next battleground for AI office suites. ② Shanghai's "Ask SO" plugged into the Qwen open platform — the first mainstream-media news-service agent and the platform's first intelligent-news ecosystem partner; Qwen App MAU has topped 300M. ③ Daiwa reiterated "buy" on Alibaba with a HK$180 target (Sept 29): disclosed Alibaba plans to scale Qwen models to 5–10T parameters, 20GW+ of global Alibaba Cloud data-center capacity by 2032, Zhenwu V900 chip in 2027 Q1, and ~¥2.1T cumulative capex 2025–2032 (analyst figures, single source).

Ahead of the National Day holiday, announced deep partnerships with several top service providers, adding flight/train tickets, ride-hailing and navigation — completing the "travel with Doubao" entry point; QuestMobile: Doubao/Qwen/DeepSeek MAU 382M/167M/130M, the top tier of China's consumer AI apps.

Meta Muse's breakout (2.5M+ downloads in 13 days; Apptopia estimates 642K US mobile DAU — nearly 3x ChatGPT's early 231K; Meta +11.43% on Sept 21, +$190B market cap) lit up the race — ① Tencent's internal Handy Bot personal agent surfaced: planned as a standalone app; its WeChat service account went live then was deregistered (single source). ② Doubao has been testing a personal-assistant product codenamed Spell since April (Sept 29 sources, single source); reports say the Doubao team worked through the Mid-Autumn break to rebuild an experience version benchmarking Muse by Sept 30 (single source). ③ Alibaba Cloud already announced personal-agent plans (Sept 22, previously covered). Debate: Muse's market case proved demand and distribution, but failure rates rise as task chains deepen and multimodality is weak; Amazon has blocked Muse from shopping on its platform since Sept 20 (unauthorized AI agents violate its terms); one privacy controversy over Muse sending a user's home address to a Marketplace buyer.

Manus 2.0's new Cascade architecture (official claims: −23.2% tokens, −28.2% task time, −32% cost; not independently verified), Manus Studio (video editing/Game Dev), Cloud Computer, event-triggered Automations, and Cue — a standalone personal-agent app where each agent gets its own email/phone/wallet/computer, invite-code beta; cryptobriefing says it's raising $500M at a ~$4B valuation (single source, unverified).

Open-sourced JEV-27B (Apache 2.0, built on frozen Qwen3.8-27B): single-forward-pass calibrated probabilities for yes/no/choice/scoring while also doing regular generative inference; runs on a single Nvidia B200 (aggregator relaying Hugging Face, single source).

No significant additions in this window.

A critical read of the White House accord — the pact writes "external evaluation, board accountability" into a joint commitment among competitors, but defines no evidence standards, auditor selection, conflicts of interest, reporting deadlines, public disclosure or consequences for violations; "its value will be measured by the records outsiders get when controls fail, not the unity of the signing photo."

Full replay of Mindgard's jailbreak of Kimi K2.6/K3 Swarm — the core story isn't the jailbreak itself but Moonshot's two-month silence after notification and response only after BBC involvement; also flags governance concerns against the backdrop of Alibaba/Tencent pushing it to a $35B valuation within a year.

An audit of the "sovereign European AI" narrative after the €3B raise — the capital map after Samsung's lead and BlackRock/Advent/Nvidia's entry, plus Mensch's "safety debate masks negligence" line vs the US "slow down" camp.

OpenAI's new $30B round / $1.4T pre-money valuation: still Bloomberg single-source (relayed by ad-hoc-news, Sept 30); OpenAI officially acknowledges only "early talks".

SoftBank's final $10B tranche due Oct 1 (completing the $30B commitment): ad-hoc-news single source; no Reuters/Bloomberg original seen.

SpaceXAI's unified Grok + X four-tier subscription (free / Lite $8/mo / Ultra $100/mo): Bloomberg exclusive on internal documents, original paywalled, secondhand only.

Meta Muse "5M downloads, faster than ChatGPT/Grok/Claude": Sensor Tower data via Cointelegraph→hokanews relay; original not seen.

Mistral CEO's "safety debate masks negligence" remarks: GuruFocus/TradingView relays; original interview source not found.

Kimi K3 / GLM 5.3 entering OpenAI's B2B marketplace: ITHome single source; no Baseten/OpenAI English announcement found.

Mindgard jailbreak of Kimi: no BBC original link found; RocketNews/Startup Fortune secondhand reporting.

Bernstein's "Huawei takes 50% of China's AI chip market by end-2026, Nvidia falls to 8%": industry-site relay of a research note, unverified.

Manus raising $500M at a $4B valuation: cryptobriefing single source.

Meta claiming R&D tax credits for AI data centers (NYT): only Cointime relay; NYT original not read.

Tencent Handy Bot, Doubao Spell beta, Doubao holiday rebuild: all Chinese-media reports citing insiders, single source.

Google open-sourcing RRSI (regularized recursive self-improvement) with Claude Opus 4.8 benchmark gains: aggregator relaying a GitHub announcement, single source.