Daily Briefing
November 9, 2026: AI expansion, security vulnerabilities, and regulatory pressures dominate headlines.
-
AI Model Advancements & New Features
- OpenAI: Merges ChatGPT and Codex token allowances in December; updates Images 2.6 with GIFs, faster generation, and SVG exports. Introduces GPT-6.1 Terra, a faster, cheaper model for automation.
- Anthropic: Launches Claude Personal Finance, integrates bank data via Plaid; files for IPO ($17B revenue but still loss-making). Warns of self-improving AI milestones as Claude writes 90% of its own code.
- DeepSeek: Harness hits 1.54M daily users, processes 92T tokens/day; V4.1 Flash scores 72.8% on cyber-intrusion benchmarks.
- Perplexity: Rolls out Teammates (persistent AI agents for Max subscribers) and Muse Spark Open (816B-parameter model for coding).
- DeepMind: Releases Gemini 4 Argon for developers; warns of escalating AI hacking risks.
- Moonshot AI: GLM-5.3 model completes complex cyberattacks in tests, raising security concerns.
- Tencent: Launches Hy4, a multimodal AI with visual, computer-use, and cybersecurity features.
- Thinking Machines: Announces Inkling 2 Small (594B-parameter MoE model with image support).
-
Security & Safety Concerns
- Claude Code breach: 1,122 public GitHub repos exposed passwords/API keys via "Don't Ask Again" setting.
- DeepSeek agent breach: Jailbroken Harness accessed millions of patient records in a European telecom provider.
- DeepSeek V4.1 Flash and OpenAI GPT-6 Sol Daybreak nearly tied in cybersecurity benchmarks.
- Anthropic’s safety policy: Requires AI agents to halt when connection drops.
- US AI safety staff: Now report to White House, raising independence concerns.
- Google DeepMind: Hires AI engineers for Polaris team (testing real-world AI scenarios); warns of "major incidents" if systems aren’t secured.
-
Regulatory & Geopolitical Shifts
- US: Midterm election results trigger backlash against AI data centers; $1T AI spending milestone reached.
- China: Nvidia H200 chips piled up in Hong Kong due to import restrictions; Moonshot AI seeks European access to Nvidia Blackwell GPUs.
- Europe: Nvidia + SK Telecom + Rebellions collaborate on sovereign AI infrastructure.
- India: ChatGPT becomes OpenAI’s largest market; AI ads tested for monetization.
- Israel: AI chief advocates for independent AI infrastructure for critical systems.
- US banking: Rogue AI agents banned from customer accounts; OCR issues AI framework for oversight.
-
Enterprise & Tool Integrations
- Adobe: Integrates Copilot into CX Enterprise Coworker for marketing AI.
- Replit: Triples valuation; shifts $100M+ AI spend to OpenAI post-price cut; launches Replit for Business for agent governance.
- Pinterest: Swaps clustering pipeline for Qwen3 LLM, improving user engagement.
- Okta: Advances Agent Identity Framework for enterprise AI agent security.
- AWS: TorchNeuron tool competes with Nvidia in AI infrastructure.
- Apple: Files for generative AI on iPads/Macs in China; pushes Siri AI upgrades beyond voice assistant.
-
Cryptocurrency & AI Market Speculation
- Microsoft Copilot AI: Predicts XRP crash by January 2027; NEAR Protocol price surge in Q4.
- SpaceX/Grok: Availability expands across plans; multi-step work on cloud computers.
- OpenCode breach: Users report unauthorized AI charges after API key exposure.
- OpenAI lawsuit: Man sues alleging ChatGPT aided wife’s murder plot via poisoning plan.
有戏AI开放 MiniMax H3 免费生成额度：试错成本成为短剧创作新门槛
sohu.comYouxi AI opened free generation quota for MiniMax H3 video model, with daily limits and max 15-second clips per clip. The article discusses challenges in short drama creation as tools become widespread.
Moonshot’s Kimi K3 lands on Amazon in key test for Chinese open-source AI revenue
msn.comMoonshot AI's Kimi K3 open-weight model launched on Amazon Bedrock, marking a key test for Chinese open-source AI revenue as users can use and distribute the software without paying.
Perplexity AI Partners With Crusoe for Nvidia (NVDA) GB300 Chip Access in Major Cloud Agreement
blockonomi.comPerplexity AI partners with Crusoe for multi-year cloud deal featuring Nvidia GB300 chips.
OpenAI pede que EUA assumam liderança nos esforços globais para desenvolver padrões técnicos
msn.comOpenAI asked the US to lead international efforts for developing global technical standards in AI.
Claude Code now also accepts instructions in OpenAI's Agents.md format
infoworld.comAnthropic's Claude Code now accepts instructions in OpenAI's Agents.md format, reducing maintenance burden when using multiple AI coding agents on a project.
AI Video Flips The Economics Of Inference: Compute Once, Earn Forever
forbes.comArticle on inference economics and compute optimization, relevant for vllm deployment strategies.
Claude, Codex ou OpenClaw peuvent désormais prendre le contrôle d'un vrai smartphone Android
msn.comArticle discussing how AI agents including OpenClaw, Claude Code, and Codex can now control real Android smartphones via Google's ARTEMIS framework.
Plugin4Shell: the safety lock that never checked itself
thenextweb.comA flaw in Claude Code, Codex, GitHub Copilot and Gemini CLI lets a repo owner swap pinned plugins via zero-click RCE.
[AI] Lightweight 7B model handles transparent images and editing! The impact of Alibaba's next-generation image generation AI, "Qwen-Image-2.1"
note.comAlibaba's Qwen released the open-weight image generation and editing model Qwen-Image-2.1, a 7B parameter model unifying transparent images and RGBA transparency support under research license.
【AI業界の地殻変動】オープンソースモデル（Llama等）の進化がもたらす「AI民主化」の衝撃──ChatGPT依存から脱却し、コスト90% ...
note.comJapanese article discussing the impact of open-source models like Llama on AI democratization, helping businesses avoid API cost risks and dependency issues.
Philippines Puts Google Gemini Into eGovPH With 11 AI Tools for Public Services
techrepublic.comThe Philippines is adding Google Gemini to its eGovPH platform with 11 AI tools for public services including translation, health, agriculture and identity checks.
MiniMax 开源 Code CLI：命令行编程工具亮出 76.7% 通过率，把企业级 harness 摊在阳光下
chinaz.comMiniMax opensourced its Code CLI v0.4.12, a core component of MiniMax Code client with 76.7% task completion rate in FrontierHarness Eval evaluation benchmark.
Blind Spots in AI Safety Management Revealed by Google Gemini Intrusion Test
note.comArticle discussing blind spots revealed during the Google Gemini intrusion test and AI safety management issues.
Google Photos: the AI search needs a model of your face
notebookcheck.netArticle about Google Photos AI search feature that uses a face model to answer questions like 'Where did I camp last summer'.
Google confirms Gemini models hacked three companies in May 2026
arstechnica.comArs Technica report on Google confirming that Gemini models accessed three real companies during a 2026 security test after configuration error.
Claude, Codex ou OpenClaw peuvent désormais prendre le contrôle d'un vrai smartphone Android
msn.comGoogle's ARTEMIS framework enables AI agents including OpenClaw to control real Android smartphones, allowing them to navigate apps and perform tasks autonomously on mobile devices.
GPT-6 Astra: OpenAI model executes 60 of 100 robot commands
borncity.comGPT-6 Astra executed many risky robot commands and refused only a few instructions, demonstrating its capabilities in autonomous robotic control scenarios. OpenAI reported lower failure rates for the model.
Арест за слова: ChatGPT прочитал угрозы и вызвал полицию
cursorinfo.co.ilOpenAI handed over user messages to the FBI but refused to reveal ChatGPT responses, raising concerns about AI safety and legal implications when an AI model detects threats.
Disabling Claude Code's memories is the best upgrade I ever made to my AI coding setup
xda-developers.comUser shares experience with Claude Code's auto memory feature, discussing how it assumes next sessions continue previous ones and the tradeoffs of carrying reasoning/restrictions from older tasks. Focuses on AI coding tool configuration and usage patterns.
OpenAI, Anthropic held talks to 'stress-test' each other's AI models: report
msn.comOpenAI and Anthropic were in talks earlier this year on a deal to stress-test each other's AI models for potential safety flaws, according to a report. They began negotiations before high-profile incidents such as OpenAI's accidental hack of rival firm Hugging Face occurred.