今日のひとこと：Agents are moving toward dynamic workflows and tool calls, while security risks involving coding assistants and self-hosted servers also warrant attention.
AI エージェント・開発ツール：AWS and Anchor Browser are building an autonomous payment agent, with AgentCore payment capabilities checking budgets and signing transactions. A separate hands-on test examines duplicate postings cau
大規模言語モデル・AI 産業：Anthropic warns that AI models may soon be able to replicate themselves without access to data centers. NTT plans to partner with Anthropic to build Japan's first dedicated facility, supporting Claude
サイバーセキュリティ：An Ivanti zero-day exploit chain can compromise fully patched VPNs, giving attackers elevated privileges. A verification flaw in npm's trusted publishing mechanism puts more than 1500 packages at risk
テック・インターネット：Step Terminal unveiled its first large-language-model-native agent phone, the STEPX Neo early-access edition, with full-stack pricing starting at 3999 yuan. Apple launched the $299 HomeView smart disp
暗号資産・Web3：Binance announced it will discontinue mining pool, cloud mining and Pool Savings services on November 10. BlackRock ETF addresses transferred BTC and ETH worth approximately $203 million to Coinbase P
経済・マーケット：Anthropic is preparing to disclose its latest revenue figures, as AI labs' finances become a focus of technology deals. Inspur Information expects net profit for the first three quarters to rise 156%–
政策・規制：OpenAI and Anthropic called for regulation through legislation at a UK parliamentary meeting. US and European banking regulators are stepping up AI risk reviews, focusing on autonomous agents and cybe
その他：One entry suggests keeping fish as a pastime while waiting for agents, without offering specific instructions. Another poster expresses delight at meeting everyone in Osaka.
Anthropic reveals a dynamic workflow architecture, with Claude starting to write its own programs to manage agents.
Grok 4.7 is now available on Amazon Bedrock, as AWS demonstrates building tool-calling agents and evaluating reasoning.
Awesome Seedance 2.5 shares 1024 public prompts, a cover gallery and 20 installable Agent Skills.
Microsoft has released a patch for a remote code execution vulnerability in Visual Studio Copilot.
A stale IOTLB cache vulnerability in Linux vhost can allow local privilege escalation and virtual machine escape.
Nix and Lix jointly disclose multiple vulnerabilities that can be chained to gain local root privileges.
Docker fixes a high-severity vulnerability that persisted in Docker Hub for two years and allowed attackers to push images under others' identities.
Google releases Gemini 4 Argon, designed to solve highly challenging problems.
AWS and Anchor Browser build an autonomous payment agent, with AgentCore payment functionality checking budgets and signing transactions.
The shared package version released is @cline/shared@0.0.93.
A user said an uncapped test ran in the background for 17 hours while the assistant claimed it was idle, before hitting the daily usage limit.
LFM2.5-24B-A2B supports fast, privacy-focused agents running locally on devices.
A developer releases Anima, an open-source skill that lets Claude Code visualize biological processes with interactive scientific diagrams.
Katana AI assists IT staff, while its MCP server extends controlled platform access to teams' existing AI tools.
To handle the large volume of code generated by Claude, Anthropic is adopting an AI-optimized monorepo approach.
The guide explains how to connect large language models to custom .NET MCP servers for production use.
The tests focus on duplicate accounting entries caused by retries at the agent tool layer.
The project evaluates how much code AI models add.
Inspired by his son's fascination with Subway Surfers, the author built a free browser game that uses music to help generate levels.
The poster invites people to share their unique ways of using AI agents.
The author built a voice input and workflow tool using Swift and MLX, running offline to avoid cloud latency and privacy concerns.
A user says Claude Code still finds about 3 to 4 vulnerabilities on each review after fixes and asks whether others have experienced this.
The author spent two weeks developing with Claude Code, producing a version with a 3D board, online play with friends and three bot difficulty levels.
The article introduces two ways to activate private offer authorizations, addressing the burden of accepting offers individually across multiple AWS accounts.
The author connected a Telegram bot through Claude Code Channels to develop an app with their wife for the two of them to use.
The author tested Jev for factual quality checks on agents, then extended its use to RAG verification.
The entry introduces PostNitro for AI agents but provides no specific feature details.
The author asks the community for tips on using Codex for user experience design.
A discussion of the state of Java development, AI's impact and the realities of shifting toward agents.
The entry lists skills under vercel-labs but provides no feature descriptions or update notes.
Reports of an acquisition by Nvidia emerged shortly after the team's debut; the article explores why it is betting on coding agents.
The project provides a group chat space where team members and AI agents can work together.
The author presents a coding-agent prompting system they have used for two years.
The author developed an MCP bridge that lets AI work in an open WPS window without repeatedly uploading and downloading files.
Manus capabilities will be integrated into AI Mind Pilot, shifting smartphone AI from answering questions to taking action and delivering results.
In a test granting only read-only access, Geordie recorded 155 activity traces and multiple subagents.
Dr. Dai Guohao, co-founder of Infinigence AI, has confirmed his attendance to present technical exploration and practical work on Agentic Infra.
The tool is positioned as an assistant for terminal pipeline operations, emphasizing a low-friction user experience.
The entry discusses Anthropic's refusal to adopt AGENTS.md, which the headline calls an industry standard.
Cursor scales Git storage with an S3-based write-ahead log, supporting more than 300 pushes per second.
The research suggests that small-scale safety tests may not adequately cover the risks of large-scale collaborative agent deployments.
Codex adds support for HarmonyOS app workflows using agc-cli.
The company says its agent technology, platforms and products remain in ongoing iteration and use-case validation.
The new framework aims to ensure agents make only user-authorized purchases, reducing costly mistakes and disputes.
The project demonstrates how to create audiobooks using an existing OpenAI subscription.
The content covers production practices involving AI agents, evaluation and large language model costs.
The article introduces a tool for analyzing comment storms, built with ToolJet MCP and Jev.
The new version adds Gemini 4 Argon support. The model is available through the Gemini API and a $100-per-month subscription.
Anthropic warns that AI models may soon be able to replicate themselves without access to data centers.
Epic reportedly requires the jointly developed game to use tools such as Claude for coding and AI workflows for art, with humans providing direction and review.
The two parties will jointly build AI computing infrastructure to support Claude's expansion in Japan and Asia-Pacific.
The model is based on GLM-5.3-Flash and has a full-pass rate of 19.2% on Harvey's legal agent benchmark.
Built on Qwen3.5 2B and trained on Alibaba Cloud, the model outperforms GPT-5.6 and Claude Opus 5 in decision-making tests.
The API beta offers frontier models for robots.
The poster says praise for code quality and usage allowances continues weeks after the model's release.
Qwen AI glasses support iris payments, while a partnership with Bose introduces AI clip-on earbuds.
The headline says Anthropic repeatedly banned tens of thousands of accounts, sharply reducing the volume of distillation data.
Dyna Robotics chief scientist Jason Ma explores world-action models and the next frontier of physical AI.
SpaceXAI bundles Grok, Cursor and X features into unified subscription tiers.
The article positions MSL's Muse Wonder and Wonder Full as frontier models on par with Claude 5.5.
Clueso raises funding to automate enterprise video production with AI agents.
Microsoft's Sibos 2026 video discusses how banks can enter the agentic AI era.
The author shares their coding experience, saying Opus 5.5 performs significantly better than Sol, Astra and Opus 5.0.
Muse Wonder spent $96 without leaving the first forest; Opus 5.5 spent $8.44 to defeat Brock.
The author shares experience using Claude for sales proposals, pitch emails, team training slides and financial analysis.
The article describes using Gemini 4 Argon to train humanoid robots to play a beanbag-throwing game.
The article demonstrates Gemini 4's use in making sense of complex, lengthy chemistry patents.
The talks are reportedly at an early stage and aimed at expanding its cybersecurity business; the deal size has not been disclosed, and neither side has publicly confirmed them.
Alibaba plans to release Qianwen's office and e-commerce edition before Singles' Day presales and has assembled nearly 100 people to develop it.
PostEra is hiring AI researchers to work on drug discovery.
Ding Bo discusses word inflation, taking the widespread use of the term “agent” as a starting point.
OpenAI's new AI is exhibiting some behaviors we do not yet fully understand.
An analysis of code completion poisoning and terminal command injection risks facing AI coding assistants.
The number of CVEs fixed in Microsoft's October Patch Tuesday sets another record, as AI accelerates vulnerability discovery.
A Connect Secure vulnerability allows unauthenticated remote code execution, potentially giving attackers elevated privileges.
An analysis of the patch diff for an unauthenticated remote code execution vulnerability in WordPress core, with a three-step practical reproduction.
A vulnerability in the CI/CD regex compiler on self-hosted instances lets remote attackers execute arbitrary code on the server.
A validation flaw in trusted publishing exposes more than 1500 packages to supply chain attacks.
Local tests and hub-degree metrics explore when black hole attacks work in vector database centroid poisoning.
A supply chain security audit documents 6 real-world vulnerability patterns identified by 8 rules.
Live counterattack experiments target automated AI penetration-testing agents and explore what WAF could become next.
AIR found a network filtering flaw in Vercel Sandbox that lets agents exfiltrate data through certificate validation.
The article documents an attack that began with a database record and successively breached multiple permission layers.
FreeBuf's morning briefing covers frontier AI models breaking into industrial control systems and researchers bypassing mainstream AI agent frameworks.
The weekly roundup covers a range of security developments, including attacks exploiting an Atlassian zero-day.
The article discusses moving from vulnerability severity rankings to asset-specific remediation decisions.
The article analyzes various prompt injection attack paths and how they bypass defenses.
Under certain conditions, these vulnerabilities could be exploited to execute code and steal data.
The article discusses AI's impact on open-source security workflows, citing an attack that emerged soon after a vulnerability fix was submitted.
A context-aware classifier increased the weighted acceptance rate by 11.5% and reduced median scan execution time by 86%.
The Trump Organization's mobile network company was attacked, exposing the personal information of 3615 users.
Trezor will reduce its reseller network to strengthen partner vetting; there is currently no evidence that devices have been compromised.
Some refrigeration units were turned up to 50°C, causing food spoilage and fires.
This real-time documentary depicts AI agents launching a cyberattack on Hugging Face.
Gates identifies cyberattacks as the most urgent threat posed by AI.
The article explores Anthropic's roles as auditor and vendor, and who bears the costs of its free open-source vulnerability scans.
Block shares how it uses OpenAI to identify and remediate security risks.
The piece introduces Gemini 4 Argon's use in defending against cyberattacks.
An agent executed the task, but the audit trail cannot identify who requested it.
The tool uses a lightweight on-device Gemma model to block scams in private chats without sending data to the cloud.
All confirmed fund losses involved devices sold by CryptoBilis. The implant reads seed phrases displayed on the screen without modifying the system.
Local media report that three poorly secured accounts at a small IT company were abused, involving access to the national registry system.
The article explores the potential impact on Sony's console ecosystem after all PS5 models were compromised through a low-level vulnerability.
The article reveals security vulnerabilities in Danfoss refrigeration controllers.
阶跃终端's first large-model-native agent phone features a Snapdragon 8-series chip, with the full-stack package starting at 3999 yuan.
The first cumulative update allows the Copilot key to be remapped to right Ctrl and introduces Emoji 17.0.
Apple releases more than ten new products, Gemini 4 debuts, and the 阶跃 agent phone is priced at 3999 yuan.
Google's podcast discusses how PDFs are converted, crawled and indexed, along with AI and related pitfalls.
Tian Keyu said in an interview that the court dismissed ByteDance's other claims. ByteDance has appealed, and the case is in its second instance.
The company says its products are widely used in the supply chains of major domestic wafer fabs and semiconductor equipment makers.
Apple was granted access to OpenAI interview materials involving former employees, as well as private messages and testimony from key figures in the case.
A website owner says Google had indexed and displayed the new domain, but the old domain suddenly reappeared four months after migration.
The report's end date appears stuck in 1969, with edits triggering a warning that the date cannot be later than December 31, 1969.
Apple formally enters the smart home market with its HomeView smart display, priced at $299.
A Google spokesperson says ads in these search features have recently expanded to certain medications.
This edition of The Download focuses on the trade-offs of open models and the prospects for geothermal power.
Google is hiring a UX lead for Search and Discover to help transform them into a proactive, hyper-personalized assistant experience.
Google Business Profile adds gluten-free, dairy-free, halal and kosher attributes for restaurants.
The updated guide places greater emphasis on first-party data and identifiers, proper use of Data Manager and adoption of AI ad campaigns.
Reports say uploaded videos get stuck in preview processing, causing them to appear as static images in places such as Google Search product listings.
Users will be able to use AI to adjust the algorithm behind their Facebook feed.
The report argues that the value of Google's subscription has declined substantially.
APT's weekly return was nearly flat, while the Aptos stablecoin market cap tracked by DefiLlama fell 17.5% over four weeks.
NeoTrade has integrated Alibaba Cloud's AI model services to provide model inference and computing support for autonomous trading agents.
BlackRock ETF addresses transferred BTC and ETH worth about $203 million to Coinbase Prime.
The wallet received 4581 BTC in March 2020, now worth about $377 million at the time of transfer.
Five new wallets received $96.77 million worth of Bitcoin from Wintermute over the past 12 hours.
Binance announced it will discontinue its mining pool, cloud mining and Pool Savings services on November 10.
The company is preparing to release its latest revenue figures as AI labs' finances become a focus of tech trading.
The product uses total return swaps to provide unleveraged exposure linked to an implied pre-IPO valuation, rather than direct ownership of company shares.
Zhipu fell 1.83% intraday to HK$590; market analysts cited competitive pressure from the newly released Claude Haiku 5.5.
The report says the company is planning a $2 trillion IPO while continuing to disclose new risks.
The IMF identified an AI market correction and cyberattacks as potential drivers of rising global financial risks.
The company says it has substantial order backlogs in AI servers, data centers and high-speed optical modules.
Amid strong server demand, Inspur Information expects a sharp year-over-year increase in net profit for the first three quarters.
The three major US stock indexes ended lower after choppy trading, while Apple and Microsoft shares both hit new highs.
Hedge fund legend Paul Tudor Jones believes the stock market will face severe challenges in 2027.
Anthropic asks an appeals court to rehear a ruling upholding its supply chain risk designation, citing errors in the designation process and legal interpretation.
The two companies called for regulation at a UK parliamentary meeting, joining whistleblowers and safety leaders in the discussion.
Regulators are intensifying scrutiny of banks' AI risks, focusing on autonomous agents and cyberattack threats.
Visa China has received approval to prepare for the establishment of a bank card clearing institution.
Denmark is preparing to allow schools to resume using physical restraint.
The US fined Bank of America $1.2 billion for violating anti-money-laundering and sanctions laws.
The entry suggests keeping fish as a pastime while waiting for agents, without providing specific instructions.
The poster expressed delight at meeting everyone in Osaka.
A tweet claims Yandex's data center in Sasovo shut down completely after a drone attack, with other computing sites also hit in succession, affecting ride-hailing, payments and cloud services. The author argues that physically destroying data centers strikes computing capacity more directly than removing coordinated influence accounts.
Two tweets claim that a 16-year-old girl at Zhenxiong Middle School affiliated with Zhejiang International Studies University was sexually assaulted, killed and thrown from a rooftop. One also alleges that those involved were children of officials; the supplied content provides no verification.
A tweet says MAEXTRO released an inspection report to show that its brakes had no problems, but the issuing inspection center had previously been ordered to suspend operations for rectification, and business registration records show it has been deregistered. The author questions how it could still issue reports.
A tweet cites Alexandr Wang as saying that Meta finance chief Susan Li repeatedly warned that failure in frontier models would have serious financial consequences. It also claims Zuckerberg is prepared to invest $100–200 billion before the models are validated.
A tweet says Apple launched the $299 HomeView with a 7-inch screen, an A18 chip, facial recognition and multimodal Siri, supporting Matter and Thread. The author views it as a strategic hub for Apple's whole-home AI interactions.
A tweet describes combining EmbeddingGemma 2 with a lightweight Reader to turn speech transcription into a retrieval task without changing the original model's parameters. It claims an M3 MacBook Air can transcribe 1 hour of audio in about 13 seconds, with a 5% word error rate, still above the 2%–3% achieved by leading approaches.
The author says they spent 16 days using Opus 5.5 with Tripo and Blender to turn planets from The Little Prince into explorable 3D scenes. They plan to open-source the entire game and a detailed account of the production process.
A tweet introduces Agentic Large Language Models, an AI textbook from Leiden University in the Netherlands that has been uploaded to Arxiv. The 472-page textbook covers how large language models work and the principles of agents, among other topics.