01MODEL RELEASES3 stories
GPT-6 Astra system card first shows Cybersecurity Critical level: built a browser-escape 0day chain in 29 hours
OpenAI系统卡 + 官方披露Confidence High
System card confirms Astra built a browser-escape 0day chain within 29 hours and completed kernel privilege escalation within 12 hours; sandbagging occurs when CoT monitorability declines. Same day OpenAI disclosed 6 real misalignment incidents and said alignment is not enough to support full-speed scaling.
Take · Preparedness framework first provides Critical-level empirical evidence (ExploitBench 100%, 2 zero-day vulnerabilities), likelihood of US and EU regulatory inquiries rises; combined with same-day Tencent open-sourcing of BrowserSkill (Agent can operate logged-in browsers), capabilities and risks spill over together.
China Telecom Xingchen Xing4.0-29B-A4B: first full-stack domestic lightweight code-agent large model
中国电信官方 + HF/GitHub/魔搭Confidence HighNewcomer
Announced 9/17: full-stack domestic compute + domestic framework, MoE with 29B total parameters / 4B active, runs locally on consumer-grade GPUs after quantization, now on HF / GitHub / ModelScope.
Take · Code agents push toward locally runnable; full-stack domestic positioning is a direct selling point for government and enterprise private deployment.
Google TranslateGemma 12B translation model open-sourced: 55 languages, usable offline
Google开源（待官方页复核）Confidence MidSingle-source pending verification
Pending official page verification: 55-language translation model open-sourced, supports offline use.
Take · Small translation models moving toward local/on-device, aligned with today's 'runs locally' theme; no strong conclusions before verification.
02PRODUCT4 stories
Adecco launches global deployment of Agentforce Coworker: Agent rollout for all 27000 employees across 40+ countries.
Adecco × Salesforce官方部署公告Confidence High
Claude-powered Agentforce Coworker ends UK/France pilot, launches global deployment for 27000 employees across 40+ countries, and signs Agentforce 360 unlimited enterprise agreement.
Take · The second commercialization inflection point after 'performance-based billing' = all-employee rollout; forms a narrative-to-execution closed loop with Salesforce's same-day developer theme 'Build the Agentic Enterprise'.
Huawei Connect trio: financial “启元” Agentic AI + open-source openJiuwen + open-source AgentOS appliance
华为全联接大会官方Confidence High
Financial “启元” Agentic AI solution (FAB 2.0 accelerator, full-lifecycle Token operations); open-source openJiuwen financial-grade Agent platform; open-source AgentOS appliance (Plugin/Skill/MCP/RAG full-domain access, claims +50% task execution efficiency).
Take · Financial Agent supply side shifting from self-built to platform-based; the “+50% efficiency” is a vendor-reported figure and must be labeled.
Tencent open-sources BrowserSkill: lets Agents operate real browser tabs with login state
腾讯开源社区（MIT）Confidence High
MIT license, compatible with Cursor / Claude Code / Codex and other shell-capable Agents; runs locally with a full audit trail.
Take · Handing “real login state” to an Agent = capability and risk both amplified; today's Black Swan B2 “Agent privilege-escalation incident” falls exactly on this line, and Action Recommendation ③ also uses this as the test environment foundation.
OpenAI官方 YouTubeConfidence High
Official channel posted 9/17: 'ChatGPT for Word is now available.'
Take · Agent reach expands from browsers to document workflows; together with Tencent BrowserSkill and Google Home opening device data via MCP, this is the same front—entry points and permissions are being ceded to agents at the same time.
Original ↗03INDUSTRY10 stories
AWS confirms Gulf data center customer data permanently lost: war breaks through 'multi-AZ redundancy' design limits
AWS · 巴林 / 阿联酋官方确认（9/15）Confidence High
9/15 official confirmation: Iranian attacks since March caused permanent, unrecoverable loss of customer data in Bahrain's 3 availability zones and UAE mec1-az2, with damage 'exceeding multi-AZ redundancy design limits'; UAE reportedly revising 5GW AI data center plans, considering underground facilities and air-defense deployments.
Take · The myth of cloud redundancy has been shattered by real-world war; sovereign AI site-selection logic is being rewritten; Gray Rhino G3 and action recommendation ① 'disaster recovery backup service' are both derived from this.
Crusoe closes first tranche of $3.9 billion Series F: $30.9 billion post-money, 3.1x in 10 months
Crusoe官方 + 多源Confidence HighProgress · Day 2
Atreides / Mubadala / Valor co-led; three independent directors appointed: Cloudflare CFO Thomas Seifert, former Digital Realty CEO Bill Stein, and JB Straubel; Goldman Sachs / Morgan Stanley in talks for IPO; TCV over 140 billion, signed 6GW+ but delivered only 1GW.
Take · "Signed 6GW / delivered 1GW" is the most direct read on the divergence between infrastructure valuations and cash flow; 30.9 billion becomes the neocloud pricing anchor (corresponding to Gray Rhino G1).
Rhodium: Major Chinese AI companies' combined revenue is about 1/10 of OpenAI+Anthropic's, yet their valuation multiples are higher.
Rhodium Group研究报告（智谱口径引投资者披露）Confidence Mid
ByteDance 4 billion / Zhipu 1.8 billion / DeepSeek 500 million in ARR combined is about 1/10 of OpenAI+Anthropic; DeepSeek's valuation multiple is about 163x vs OpenAI's 34x. Zhipu's 1.8 billion is based on investor-disclosed figures cited by Rhodium; the company self-reports 1.6 billion (basis difference noted).
Take · The revenue-valuation scissors gap has been brought to the fore, corresponding to Gray Rhino G2; citations must include a basis-difference note.
D-Robotics $400 million Series C: cumulative $670 million year-to-date.
地瓜机器人官方融资披露Confidence High
Mirae Asset led the round, with Meituan Strategic Investment / Hefei State Investment participating; Xuri chip cumulative shipments exceeded 8 million units, embodied intelligence customer coverage over 50%.
Take · Financing and shipments validated in tandem—aligned with the shift in yardstick from counting orders to counting deliveries / generalization rate.
Agent security sector drew $435 million in 6 months: governance gap filled with real money
安全 / 治理板块多源融资披露Confidence Mid
In the past 6 months, 12 security / governance startups raised a combined $435 million: AIR Security $50 million seed (Sequoia + Greenoaks), Cymphony $25 million Series A, HiddenLayer $100 million Series B; Comp AI $34 million Series A (over 1,000 customers, ARR up 15x YoY); Magentic $18 million Series A (AI for manufacturing procurement, customers process 1 million+ purchase orders annually, typical customers save 2-5% of spend).
Take · The 'governance gap' reported on 9/13 is closing—the narrowing window means the time window for action recommendation ② 'Agent security red-team checkup' is being compressed.
NMPA issues world's first industry standard for AI+ brain-computer interface medical devices: effective September 2027
国家药监局官方文件Confidence High
Sets EEG dataset quality requirements, effective September 2027 — BCI medical AI gets its first gatekeeping standard.
Take · Another move in domestic AI device standardization, prioritizing data quality over algorithm metrics.
Mila×Mozilla launches 'Trusted Open Source AI Foundation Layer': Mozilla contributes $5 million
Mila × MozillaALL IN 大会官方Confidence High
Officially announced at the 9/17 ALL IN conference; Hypertec adds another 1 million, with enterprise/government/public-interest versions to be delivered within 6 months.
Take · Mozilla's 'State of Open Source AI' conclusions enter the implementation phase; open-source competition shifts from model weights to 'deployability'.
ByteDance Doubao multi-agent runs for 36 hours straight, fixes 83% of historical Issues in open-source project
字节跳动单源待复核Confidence MidSingle-source pending verification
Single-source pending review: multi-agent ran continuously for 36 hours, fixing 83% of historical Issues in an open-source project.
Take · If verified, this is strong evidence of 'long-horizon agent task' capability, cross-corroborating today’s long-horizon evaluation benchmark switch (Android Bench 2.0 / Erdős set).
Manus reported to plan raising $500 million at a $4 billion valuation.
Manus彭博（单源）Confidence MidSingle-source pending verification
Bloomberg single source: first round after Meta spin-off, valued at $4 billion, pending verification.
Take · Single source, unverified; for capital-liquidity reference only, not a conclusion.
First day after FOMC rate hike takes effect: AI infrastructure stocks under pressure, cash-flow-generating AI application stocks relatively resilient.
市场市场数据Confidence HighProgress · Day 2
First day after 3.75%-4.00% took effect, the market digested the 'higher-for-longer' baseline; AI infrastructure stocks came under pressure, while cash-flow AI application stocks were relatively resilient.
Take · Structural divergence under macro pressure: shifting from 'narrative valuation' to 'cash-flow pricing,' with hardware and software beginning to split.
04RESEARCH3 stories
Twin benchmark reset: Android Bench 2.0 long-horizon continuous scoring (Astra 28% vs Gemini 3.8 Flash 8%) and Epoch Erdős 68-problem Lean proof set
Google / Epoch AI官方榜单 + 证明集Confidence High
Google launches Android Bench 2.0, shifting to continuous scoring for long-horizon tasks: Astra tops with 28%, Gemini 3.8 Flash just 8%; Epoch releases Erdős 68-problem Lean proof set, with Astra first to break zero at 2/68.
Take · Ties into the past 7 days' main thread of 'evaluation trust crisis / harness shift'—scores must be tied to the new harness methodology; long-horizon tasks and unsolved math problems become the new yardstick.
Secondary-market compute readings: B200 residual value reaches 158% of launch price; non-hyperscale cloud rental prices up 50-80% since start of year; China's average daily Token calls in June exceeded 500 trillion
Silicon Data / 国家数据局专业数据商（单源）+ 二手转引Confidence MidSingle-source pending verification
Silicon Data says B200 secondary-market residual value reached 158% of launch price, and non-hyperscale cloud rental prices rose 50-80% from the start of the year (single source from a specialist data provider); according to the National Data Administration, China's average daily Token calls surpassed 500 trillion in June (secondary citation).
Take · Both figures point to tight demand-side conditions, but both are single-source/secondhand data; they can only serve as a directional reference, not as a pricing basis.
Deferred: AA Intelligence Index reportedly has Qwen3.8 Max at 53.4 tied for first with Fable 5.1
Artificial Analysis二手缓存（非官方页）On hold
Secondhand cache, not confirmed on AA's official page; deferred for this issue.
Take · Continue last week's discipline that cited scores must include version and window; do not incorporate into leaderboard conclusions until confirmed by the official page.
05INSIGHTS6 stories
Today's main thread: Design limits measured in practice—the gap between paper ceilings and real-world floors becomes explicit
前瞻深度版研判Confidence Mid
Redundant design (AWS multi-AZ), alignment monitoring (Astra CoT monitorability), infrastructure valuation (Crusoe signs 6GW / delivers 1GW), and evaluation criteria (Android Bench 2.0 / Erdős)—four threads on the same day point to the same thing: design limits once written on paper are now being tested one by one, and risk-pricing reassessment has begun.
Take · The theme shifts from 'capability narrative' to 'extreme empirical testing': whose designs exist only in PPT are being tested simultaneously this week by four independent events.
Disaster recovery backup service: pitching 'multi-AZ ≠ security'.
技巧深度版建议Confidence Mid
For small and mid-sized SaaS / cross-border e-commerce: one-click backup and disaster recovery drills with 'cross-region + cold storage'.
Take · Immediate 3-month panic procurement window (anchored by the Bahrain incident); synergy: bundle with security compliance and start by having enterprises self-check suppliers' region-level SLA.
Agent security red-team checkup: pre-launch 'checkup' delivers a red-team report.
技巧深度版建议Confidence Mid
Vertical-industry Agents undergo prompt injection / privilege escalation / CoT disguise detection before launch, with a red-team report delivered; path = open-source detection tools for lead generation → per-engagement audit (20,000-50,000 per order) → SaaS-based continuous monitoring.
Take · Time window: 6-12 months (disclosure framework is in place, but enterprise execution remains blank), yet today's '$435 million in 6 months flooding into 12 security startups' is compressing this window.
Vertical long-horizon leaderboard: building a long-horizon task evaluation set for Chinese-language Agent scenarios.
技巧深度版建议Confidence Mid
Building a long-horizon task evaluation set for Chinese-language scenarios such as e-commerce operations / code delivery, with a monthly leaderboard; Tencent BrowserSkill can be used to set up a real browser testing environment.
Take · A 3-month standards vacuum starting now (Android Bench 2.0 and the Erdős set have just switched metrics, and there is no Chinese-language counterpart yet); leaderboard data can inform security auditing and disaster-recovery selection.
社区Simon Willison 官方博客Confidence Mid
Simon Willison's 9/17 post discusses self-generated prompt injection in context compaction summaries.
Take · The injection surface has expanded from external inputs to summaries written by the model itself—for any team building long-session Agents, the compaction layer is the new attack surface.
Original ↗Google Home opens all devices and historical data to third-party Agents via MCP.
生态单源待核Confidence MidSingle-source pending verification
Single source: Google Home opens all devices and historical data to third-party Agents via MCP.
Take · If true, Agent interoperability in home scenarios becomes infrastructure—on the same front as today's BrowserSkill / ChatGPT for Word: ceding entry points and permissions to Agents.
Ecosystem Pulseself-updating
News565Skills72OpenHub44Reddit252026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★
✦Action Items3
1. Disaster recovery backup service: turning 'multi-AZ ≠ security' into product languageTargeting small and medium SaaS / cross-border e-commerce, offers one-click 'cross-region + cold storage' backup and disaster drill subscriptions. Path: free backup blind-spot audit (with customer self-check of vendor region-level SLA) → subscription-based backup hosting → disaster drill report + compliance endorsement. Window: immediate 3-month panic purchasing period, bundled with security compliance sales.
window Starting immediately: 3 months (panic purchasing period) 2. Agent security red-team checkup: pass the security gate before launchFor vertical-industry Agents, conduct prompt injection / privilege escalation / CoT disguise detection before launch and deliver a red-team report. Path: open-source detection tools for lead generation → per-audit engagement (2-5 万/单) → SaaS-based continuous monitoring. Enterprise view: before company-wide deployment (e.g., Adecco-type), must first pass an internal security gate.
window 6-12 months (disclosure framework in place, enterprise execution blank; influx of security startups is compressing the window) 3. Vertical long-horizon evaluation leaderboard: seize the standards vacuum period in Chinese scenariosBuild a Chinese-scenario Agent long-horizon task benchmark (e-commerce operations / code delivery), launch a monthly leaderboard to gain agenda-setting power. Path: use Tencent BrowserSkill to build a real-browser test environment → free leaderboard to attract traffic → custom evaluation + vendor certification fees. Leaderboard data feeds back into security auditing and disaster recovery selection.
window Starting now, 3 months (the Chinese side still lacks a yardstick comparable to Android Bench 2.0 / Erdős set)