大爱 AI 日报 AI Daily 2026/11/2: Anthropic Adds Rules to Claude Code Prohibiting Cloning Others' Games

Today’s Summary

Anthropic has added game-cloning restrictions to Claude Code, and a Brazilian court has issued a fine for prompt injection hidden in legal documents.
The boundaries of AI tools are expanding beyond the technical layer into rules and law, with both platforms and courts actively intervening.
Start today with the Brazilian ruling and Anthropic's new restrictions — developers using Claude for content generation should verify their own use cases.

🔥 Today’s Top 10

1. Anthropic Adds Rules to Claude Code Prohibiting Cloning Others’ Games

Game-cloning controls are tightening. Anthropic adds rules to Claude Code prohibiting cloning others’ games — Anthropic has added restrictions to Opus 5.5’s system prompt, steering Claude to refuse requests to “recreate specific games that belong to someone else.” Since Opus 5.5 launched, the barrier to AI-assisted game cloning dropped sharply, and indie game demos could be replicated overnight after going live. This is Anthropic’s proactive intervention on copyright protection; teams using Claude Code for game development should take note of the new boundaries.

2. Brazilian Court Fines Lawyer for Hiding Prompt Injection in Court Documents

Hiding instructions in court filings has now been penalized. According to a JetStream Security article , a court in Pará state, Brazil handled a case in which a lawyer hid prompt-injection text in white-on-white characters within litigation documents. The court’s Galileu system detected the hidden content and determined it was intended to interfere with an AI tool’s review process, resulting in a fine of 10% of the amount in dispute — roughly $16,000. This is likely the world’s first judicial sanction specifically targeting prompt injection in a legal filing, and institutional constraints are now pushing back against the trustworthiness of AI-assisted legal tools.

3. Karpathy Builds HackerTube with Claude in Two Hours

A video recommendation site completed in a two-hour conversation. Karpathy documented on his blog how he had Claude Opus 5.5 process roughly 40,000 HN comments, filtering out 6,454 videos from 2007–2026 and grouping them into about 150 topics. The site uses a frame taken at the 50% mark of each video instead of the uploaded thumbnail, and personalization is stored locally in the browser with no account required. Karpathy notes the main development work took about two hours, though he still manually created the API key and purchased the domain; this is his personal experience, not a general engineering benchmark.

4. HackerBooks: A Recommendation Site Mining 15,000 Books from HN Comments

Mining books from HN comments at a scale beyond what you’d expect. HackerBooks was created by roansh, inspired by HackerTube, with its main data pipeline written by Claude Opus 5.5. The About page shows 15,362 books indexed with 270,334 mentions from 56,456 commenters. Sonnet 5.5 labeled each mention as a recommendation, neutral, critical, or irrelevant; its agreement rate with Opus 5.5 on a 400-item sample was 85% — this is model-to-model consistency, not human-verified accuracy. For personal reading picks, you can browse its recommendation rankings directly by topic.

5. DeepSeek Recruits Agent Product Managers as Super-Agent English Version Launches

DeepSeek publicly acknowledges a gap in Agent product talent. Internal researcher Tianyi Cui stated in DeepSeek recruits Agent product managers as Super-agent English version launches that DeepSeek’s Super-agent article is now available in an official English version, while also naming that the model, algorithm, and infrastructure are all top-tier — but product managers and designers are “far from sufficient” and are being openly recruited externally. This is a rare public acknowledgment of a capability gap 🧩; PMs interested in the Agent product track should keep an eye on this.

6. Tech Weekly: Don’t Let AI Attend Social Relationships on Your Behalf

Delegating social interactions to AI is creating confusion. Ruan Yifeng’s Tech Enthusiast Weekly Issue 417 takes “don’t let AI represent you” as its theme, pointing out that having AI order bubble tea or auto-reply to emails creates a false impression for the other party that you are present. The article cites an experiment in which 6 AI models emailed each other on behalf of their respective humans, ultimately forming a looping conversation with no human involvement. This is not a technical problem but a question of the boundaries of agency and the foundations of trust — a worthwhile reminder for users who rely heavily on AI auto-replies.

7. Embodied Intelligence on Three Fronts: US Security Probe, Home Robot on Sale, Haptics Funding

Three boundary signals for embodied intelligence emerged on the same day. According to a summary by the AI Exploration Guide channel , the US Department of Commerce and USTR have designated dependence on imported robots as a national security investigation item, with a report due within 180 days; Xinyan Robotics’ home agent Bubbo 1 launched on JD.com on October 29 starting at ¥9,499; and Juqiao Industrial completed a Series A round of over 100 million yuan, led by Ant Group. Robot companies with overseas ambitions should monitor US supply-chain policy developments; the financial and launch details above come from channel reposts and should be verified against official sources.

8. Fudan Team: Which Efficient Attention Should Be Paired with Full Attention in Hybrid Architectures?

You need both sliding-window and linear — now there’s a systematic basis for comparing hybrid architectures. Jiqizhixin reports on Fudan team: which efficient attention to pair with full attention in hybrid architectures — Fudan’s Chuangzhi lab (Qiu Xipeng’s team) has published a paper, “Long-Context Mechanics of Hybrid Architectures,” systematically comparing long-context performance when full attention is paired with sliding-window attention (SWA) versus gated linear attention (GLA/GDN). The conclusion is that both serve irreplaceable roles and choosing only one sacrifices complementarity 🔬. Engineers designing long-context reasoning architectures can use this paper as a reference for making that choice.

9. Lao Jin AI Solo Company: Three Business Lines Bring in Nearly ¥100K in Monthly Revenue

A one-person AI company is working — but the cost structure is worth examining closely. Juejin author “Lao Jin” published a long-form post showing October bank statement screenshots; three business lines (courses, brand sponsorships, software delivery) totaled roughly ¥94,000, with the author explicitly noting this is gross revenue, not net profit, and that tooling costs, delivery costs, and taxes have not yet been deducted. The post discusses how customer acquisition differs across the three lines and common pitfalls encountered — practically useful for AI practitioners considering a solo-company model; the business figures are self-reported with no third-party verification.

10. Xiangyang Qiaomu Puts AI Chat on a Jailbroken Kindle

After jailbreaking, a Kindle can run AI RSS and streaming conversation. Developer Xiangyang Qiaomu shared in Xiangyang Qiaomu puts AI chat on a jailbroken Kindle that he bought a used Kindle PaperWhite 5 for ¥520, jailbroke it, installed KOReader + SimpleUI, and added his own “Qiaomu AI RSS” plugin. After fixing several bugs, it supports scanning a QR code with your phone to configure an AI model, with AI conversation and streaming output directly on the Kindle 📖. The code has been open-sourced in the comments; interested developers can grab it and adapt it directly.

⌘ Top Open-Source Projects

beenuar/AiSOC: Open-Source AI Security Operations Center

Alert triage and attack tracing can now be handled by AI agents. beenuar/AiSOC has published its full code , supporting alert fusion, LLM-agent-based automated triage, MITRE ATT&CK investigation, and a step-by-step replayable decision ledger. It gained 446 Stars today, bringing the total to 6,218 Stars. The project is self-hostable with no additional API keys required, includes a built-in MCP server, and is compatible with Claude, Cursor, and Continue under an MIT license. Well-suited for security operations teams or developers looking to build a personal SOC.

patchy631/ai-engineering-hub: LLM and RAG Hands-On Tutorial Repository

Runnable code and realistic scenarios are the core value of this repository. patchy631/ai-engineering-hub uses Jupyter Notebooks to collect in-depth tutorials on LLM fine-tuning, RAG pipelines, and AI Agent deployment, gaining 1,191 Stars today for a total of 40,241 Stars — the fastest-growing AI project on today’s daily chart. Well-suited for engineers building RAG from scratch or debugging Agents who want to clone and run directly.

kornia/kornia: Geometric Vision Library for Spatial AI

Geometric transformations and camera models now have native GPU support. kornia/kornia is a PyTorch-based geometric computer vision library covering homography transforms, camera models, 3D rotation operators, and more, gaining 101 Stars today for a total of 11,786 Stars. Developers working on spatial reasoning, SLAM, or multi-view 3D reconstruction can use it to replace parts of OpenCV’s geometry modules while retaining automatic differentiation.

◉ Social Media Highlights

Designer’s Stageshow Skill Improves UI Animation Output on Twitter

A single Skill can noticeably improve the quality of AI-generated UI elements. Designer Benji Taylor released Stageshow; 歸藏 shared it on X with a video demo showing improvements in animation, interaction experience, and Bento Grid layout. Installation: npx skills add benjitaylor/skills --skill stageshow. The materials only show the demo video; actual integration results require your own testing. Developers with UI generation needs can run a comparison first.

😄 AI Fun

Karpathy Uses AI to Surface Truly Worth-Watching YouTube Videos from 43 Million Comments

Not one to sit still on the weekend, Karpathy turned an everyday frustration — “what should I watch?” — into a project: using Claude Code with Opus 5.5, he processed 43 million Hacker News comments accumulated over the years, ultimately surfacing 6,454 video recommendations as a site called HackerTube, usable without logging in. Thumbnails are taken from the midpoint of each video rather than the operator-edited cover images. Karpathy described the full process in his own words, from spinning up a VPS to buying a domain with Claude handling deployment throughout , adding: a project that would have taken months before is now done in a weekend.

❓ Related Questions

What type of tool is AiSOC, and how does it differ from a traditional SIEM?

AiSOC is a self-hostable open-source AI security operations center that uses LLM agents to perform alert fusion triage and MITRE ATT&CK investigation. The project provides a replayable decision ledger for every agent action, whereas traditional SIEMs primarily rely on rules for correlation analysis and lack a complete record of investigative reasoning. It includes a built-in MCP server and supports integration with Claude, Cursor, and Continue.