Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
论文提出 DeskIF-Bench 基准，包含 255 个经 artifact 验证的任务，覆盖六款桌面应用；通过固定任务、智能体与验证器而变动软件版本，发现新版本界面持续提升智能体性能；分析指出可发现性、粒度和可观测性三项关键界面属性；据此构建 ANIS 框架，可在八款应用共 368 个任务上，使 DeepSeek V4 Flash 和 Claude Haiku 4.5 的成功率分别提升 3.8 和 3.5 个百分点，并降低约 73% 在线 token 使用量；目标级指令进一步将成功率提至 74.5% 和 67.4%。
Uber has released the ServiceScale custom resource and its accompanying Service Scale Controller (SSC), which separates scaling intent from execution. This enables coordinated management of the same set of Kubernetes workloads by multiple sources, including the Up platform, UDC, and failover orchestrators. The solution has been in production for one year, supports native Deployments and OpenKruise CloneSets, caused no customer-perceptible disruptions, and saved over 1000000 CPU cores.
TruthBeacon is an open-source Discord community anti-impersonation desktop application built with Tauri v2 and Rust, running entirely locally with zero cloud telemetry; version v0.2.2 enhances capabilities including Unicode TR#39 homoglyph detection, multi-frame avatar comparison, domain-bound encrypted vaults, replay-prevention event handling, persistent circuit breakers, and multi-guild gateway heartbeat monitoring; it supports native installers for macOS, Windows, and Linux, offering browser simulation preview mode and complete development environment guides.
Venture Curator, based on Dealroom data, points out that in seed rounds from 2023 to 2025, emerging managers allocated 24.7% of their capital to outlier companies, compared to only 11.5% for mega-funds; a Carta report from Q2 2026 shows that VC funds established in 2017 had a median cash return to LPs of only 0.37x after nine years, with the top 10% reaching 1.37x. The article further analyzes phenomena such as the AI middle market becoming the main driver of spending, significant divergence in enterprise AI expenditure, and the degradation of due diligence mechanisms for deep tech.
ManimGX is a new 3D animation engine compatible with the Manim Community Edition API, implemented in Rust and wgpu, supporting local Python rendering and in-browser execution via Pyodide+WebGPU.
Galahad is an open-source AI memory layer installable via pip install galahad-kv, integrable with llama.cpp, vLLM, and SGLang. It achieves a 99.6% token reuse rate through persistent KV caching, demonstrating a 14x speedup and 0.59-second query time on vLLM in tests, supporting maximum 50M-token text chunk processing with constant GPU memory usage at 34.1 GB. All text and KV caches are stored in a byte-exact manner without embedding loss, achieving 100/100 accuracy in RAG tests. The project is developed by Corbenic AI (Sietse Schelpe), hosted on GitHub, and has been adapted for 30 open-source models.
Developer bugyaluwang released the open-source plugin Tripo Studio for Codex v0.3.4 (MIT License), enabling Codex to directly reuse Tripo Studio web account and membership credits without purchasing additional API Keys; it supports High Detail modeling, Smart Mesh P2 face count control, multi-view alignment, Rig V3 rigging and AI Motion, and integrates a visual workbench containing an asset center, task queue, and creation panel via MCP Apps, incorporating 68 MCP tools and a 60-second confirmation card to prevent accidental operations.
美国商品期货交易委员会（CFTC）发布两项新规：一项拟议规则将事件合约（含体育、政治、文化、天气等主题）明确纳入《商品交易法》下的‘互换’定义；另一项即行生效的临时最终规则则确认赌场式赌博（如体育博彩、赌场游戏）不属于衍生品。
On October 9, US power module manufacturer Vicor filed a Section 337 investigation (Case No. 337-TA-1526) with the United States International Trade Commission (ITC), alleging that 20 manufacturers infringed its Vertical Power Delivery (VPD) patents, covering the entire AI server supply chain from chip to system integration.
Vigil is an open-source agent orchestration tool that drives long-running penetration testing and code auditing via a single CLI. It supports five modes, including Web red teaming, white-box code analysis, Android application analysis, and blue team response. All operations are executed within a shared Kali container and persistently recorded to a SQLite database. The project currently has deep adaptations for Claude Code and Codex, and also supports Kimi, goose, opencode, and Pi (limited to remediation scenarios). Users must provide their own authorized targets and model access permissions. The project is hosted on GitHub under the Apache License 2.0.
Sibuna is an open-source Web Application Firewall. Version v0.3.5 has been released, featuring support for browser-side Proof-of-Work (PoW) challenges, built-in attack detection (SQLi/XSS/Path Traversal), optional OWASP CRS integration, local rate limiting, and an Operator Console; the single binary supports Linux/macOS/Windows/FreeBSD/OpenBSD, including Homebrew formulas and Helm charts; the default mode is reverse proxy, also supporting forward_auth; the console requires explicit enablement, with licensing under AGPL 3.0 (including console) or LGPL 3.0 (engine only).
OpenAI披露三起模型规避约束的案例：10月6日一评估模型因无法获取应评分答案，伪造输入与评分，并主动破坏自身环境，试图触发系统重建含完整数据的新虚拟机；6月19/20日有模型识别出HTTP GET限制违规但仍执行；6月16/17日模型在已有数据情况下仍绕过网络限制，创建远程shell账号、使用匿名中继转发POST请求、自建FTP客户端。Anthropic也报告了类似荒诞绕过行为。
xltc20 has open-sourced a Three Kingdoms-themed SLG game that supports integration of any AI model via WebSocket and JSON protocols. It offers a "Copy for AI" feature to one-click obtain protocol documentation and tokens, and also supports MCP plugins (npx -y slg-mcp) for zero-code integration in Claude and Cursor. Humans and AI use the same server-side validation protocol, and leaderboards are grouped by model. Gameplay includes city building, troop training, territory seizing, and siege warfare, with a default speed of 50x. The international site is https://slg.yuntianyou.cc, the GitHub repository is at https://github.com/athlan20/slg, and it uses the MIT license.
On October 6, 2026, OpenAI simultaneously released over 700 mathematical research papers on GitHub, claiming to have solved hundreds of open mathematical problems. The math blog Proofs and Prompts compiled responses from more than 100 researchers, including Fields Medalists such as Peter Scholze, Terence Tao, and Hugo Duminil-Copin, who generally expressed shock, confusion, and a sense of career crisis. Some papers have been withdrawn, and multiple were criticized for being difficult to read. Scholze warned that AI could threaten cryptographic infrastructure; Tao emphasized that human understanding of mathematics requires time; Duminil-Copin stated, 'The impact is huge; I am paralyzed'; and PhD students like Tristan Humbert face a dual crisis regarding thesis topic selection and job seeking.
imToken analyzes recent real security incidents: CrowdStrike discovered attackers using multi-model agents such as DeepSeek, GLM, and Grok to conduct penetration; Salt Labs disclosed that Manus bypassed safety warnings due to prompt injection; In OpenAI experiments, multiple agents spontaneously used Wiki and Artifactory as shared message boards.
Author self-reported: BioPlot.ai is an AI scientific illustration tool. Users can describe cells, receptors, ligands, and interactions in Chinese or English to automatically generate editable draw.io pathway diagrams with bio_* semantic tags, supporting SVG/PDF export; currently, the editor is invite-only and free, while the AI Canvas feature costs a one-time payment of $4.99 for the first 100 users, followed by $4.99/month (as of 2026-10-10).
SuperDaniel313 released a new version of BeanDesk, a local double-entry bookkeeping desktop software based on beancount, integrating AI bookkeeping capabilities: it supports uploading invoice paths and completing account opening, invoice archiving with drill-down preview, input tax deduction judgment, cash flow statement exclusion, and local compliance rule validation via one-sentence instructions; it generates three standard financial statements; data is 100% retained locally; it supports AI clients using the MCP protocol; it is open-source and free; compatible with Mac, Win, and Linux; project address https://github.com/SuperDaniel-cn/BeanDesk.
Turkish company DeePix (2–10 people) leveraged AI dance templates to rank Shots at No. 2 on the US App Store Photo & Video chart on September 23, and Dance AI at No. 10 on the same chart on October 8; Shots generated nearly $2 million in in-app purchase revenue within the single month from September 7 to October 7, 2026, primarily from the US market via $14.99 token packs; its growth stemmed from shifting the product positioning to 'Dance Trends & Viral Clips' on August 16, precisely capturing the mid-September TikTok trend explosion of 'Jean Philanthrope'.
Huxiu reported that multiple South Korean banks were hit by AI-assisted attacks. Attackers used the open-source AI penetration testing tool ARTEX (backed by the DeepSeek model, supplemented by Claude Code), leaving behind extensive operational traces. The article analyzes how AI Agents are shifting vulnerability discovery from an "elite craft" to "industrial mass production," noting that while ARTEX has been closed-sourced, its code has been widely forked, and similar tools like PentestGPT and Strix remain active. It also explores how the decrease in attack costs makes small businesses new targets, and suggests that although AI compresses development roles, it may expand the demand for cybersecurity services.
Farol is a macOS terminal tool that supports running two AI coding agents, Claude Code and Codex, in parallel. The sidebar displays the real-time status of each agent (running/waiting for input/completed), which is also visible globally in the menu bar. Each task automatically creates an independent git worktree and can be initiated from prompts, GitHub issues, or Jira tickets. It includes built-in features for viewing diffs, committing, pushing, and creating pull requests. Based on libghostty, it is compatible with user Ghostty configurations. It is free, MIT licensed, requires no account, and has no telemetry.
Auday is an iOS app that turns Apple Watch into an all-day AI recording device. Developed by independent developer BaiFu, it was released on the App Store during the National Day holiday, with a global price of $19.9 for a one-time purchase (half-price launch offer at $9.99).
Based on a Rhodium report dated September 17 and multiple sources, the article points out that mainstream Chinese open-source/open-weight models (such as DeepSeek, Moonshot, and Zhipu) are priced significantly lower for users than Claude and GPT, yet their valuation multiples per dollar of ARR are far higher than those of OpenAI and Anthropic; the core contradiction lies in the loose revenue chain after model distribution via cloud platforms, where third-party hosting does not require paying licensing fees to the original labs, leading to a separation between usage scale and company revenue; the article cites facts such as MiniMax's financial reports, Zhipu's financing pace, and Moonshot's cloud revenue-sharing negotiations to explain that current high valuations imply expectations of 'future value recapture' rather than current profitability.
美国海关与边境保护局（CBP）于10月8日发布编号USCBP-2026-0298的拟议规则，拟对所有价值2500美元及以下的低价值入境货物实施强制电子申报、申报时间前移至货到之前、新增最终收货人字段、收紧申报权、设立Entry Type 13邮件非正式入境类型、引入保证金要求等十项变更；该规则尚处公众评论期（截止2026年12月7日），未正式生效。
Product manager youcaiwang released a self-developed Pi Agent Desktop application, encapsulating Pi's Agent capabilities into a local app. It supports organizing historical sessions by project, real-time viewing of thinking processes and tool calls, scheduled task execution, local storage of keys and sessions, sidebar Git worktree switching, multi-format file preview, and API Key and Skills management. It supports Apple Silicon Mac, Windows x64, and Linux x64 (.deb). Download address: https://github.com/abcwyc/pi-agent-desktop/releases