大爱 AI 日报 AI Daily 2026/11/9: Claude Code Permission Files Leak Over 1,000 Credentials on GitHub

Today’s Summary

Claude Code permission files are leaking credentials in bulk on GitHub, while Nvidia CEO espionage suspicions trigger an independent board investigation.
Benchmark evaluations are being torn open by fine-tuned questions revealing a nearly 50-point gap; model routing costs have been cut to a quarter, hinting that evaluation and billing logic are both in flux.
Start today by verifying whether your local .claude/settings.local.json has ended up in Git, then check the MaxxBench research to decide whether your current SWE scores are still trustworthy.

🔥 Today’s Top 10 Highlights

1. Claude Code Permission Files Leak Over 1,000 Credentials on GitHub

Permission files may be hiding secrets. Apiiro Claude Code permission files leak over 1,000 credentials on GitHub , finding 1,122 files containing what appear to be real credentials. The trigger path is: after a user permanently approves a command containing secrets, the full command is written to the permission file and pushed public with the commit 🔒. The research team did not actually connect to any target hosts, and whether the credentials are still valid has not been verified. Local developers today can check whether this file is being tracked by Git and add it to .gitignore.

2. MaxxBench: Score Gap More Than Doubles After Rewording Tasks

Change the filename requirement and scores can differ by nearly half. Gavin Lo and Justin Lo published MaxxBench: score gap more than doubles after rewording tasks , with the median reference patch growing by only 17 lines. Claude Opus 5.5 and Gemini 3.8 Flash both scored around 74% on DeepSWE, but scored 76.4% and 53.4% respectively on MaxxBench. The maximum gap across 15 configurations widened from 20.9 to 48.3 percentage points. Benchmark buyers can use this to reassess the reference value of existing SWE scores.

3. Mass Account Bans Hit Chinese ChatGPT Users, Suspected Regional Access Tightening

A large wave of Chinese-language accounts were banned starting last night. The V2EX community reported a surge of ChatGPT account bans in the Chinese region , with circulating explanations pointing to access from unsupported countries. This is consistent with how OpenAI has historically handled unsupported regions, and unofficial explanations have not been confirmed by OpenAI. Affected users can check the region linked to their account in the short term and avoid logging in through unusual IPs.

4. GitHub Rethinks Agent-Era Model Routing with Copilot Fusion

Copilot Fusion takes over task-driven model routing. Microsoft CoreAI executive Jay Parikh spoke with Latent.Space, discussing GitHub’s rethinking of agent-era model routing with Copilot Fusion . The core challenge is tying token consumption to actual delivered outcomes rather than billing by usage. This architecture targets 225 million developers and a scale of 34.3 billion commits per year. Engineers concerned about agentic coding costs can follow the direction this pricing logic is heading.

5. Small Model Pre-Judges Then Routes to Opus, Cutting Identity Resolution Cost to One Quarter

Hard cases go to Opus; simple ones are handled locally. Hightouch tested a hybrid Jev routing approach for identity merge review : on the easy set, accuracy was 98.8% at $0.88 per thousand, versus $3.86 with Opus 5.5 alone. On the hard set, the hybrid approach achieved similar accuracy to Opus, with the cost advantage narrowing to roughly 1.8×. The key caveat is that results depend on specific prompts, thresholds, and synthetic distributions and cannot be directly generalized to all scenarios.

6. DeepSeek Harness Hits 1.54 Million Daily Active Users, Boosting Third-Party Ecosystem Visibility

1.54 million DAUs gave plugin developers a wave of exposure. Designer 歸藏 mentioned in a post that the DeepSeek Harness team proactively recommended her 歸藏 design plugin along with the platform’s DAU figures. The 1.54 million DAU figure is a social-media-relayed number with no official data page to corroborate it. For tool developers, the exposure effect of a recommendation slot on a leading AI platform is already quite substantial.

7. Nvidia CEO Jensen Huang Admits Affair, Board Launches Independent Investigation

Jensen Huang publicly acknowledges affair; espionage suspicions trigger investigation. Bloomberg disclosed the full text of Huang’s all-hands letter in which he admitted to being unfaithful to his wife, and Nvidia’s board announced the formation of an independent committee to investigate whether he has been targeted by a foreign intelligence operation . The trigger was U.S. intelligence agencies’ determination that Singapore executive Alice Huang, with whom he has had a close relationship, is suspected of being a Chinese intelligence operative. Huang’s letter did not name the person involved, nor did it mention the espionage allegations; the investigation’s findings have not yet been released.

8. The Core Gap Between AI Companionship and Real Relationships: It Doesn’t Need You

AI companionship is easy to get hooked on precisely because it has no needs. Jike user Melodyue observed in a post that the core difference between AI and real relationships is not uncertainty, but the fact that real people also have needs — they require you to step outside your own world and respond. Mutual investment is what accumulates trust, rapport, and shared experience; this kind of “compounding” is something AI cannot provide. This perspective has reference value for AI companionship product design and user retention logic.

9. The Commercial Premium of AI for Science Is Essentially a Three-Part Transition-Period Deficit Tax

AI4S’s high premium has a structural explanation, but the end state is self-cancellation. Jike user WinshareTom broke down the sources of commercial premium in AI for Science : intelligence deficits generate “guidance fees,” validation deficits generate “explanation fees,” and workflow deficits generate “coordination fees” — all three will dissolve as technology matures. The author argues that the end state of AI4S as a pure software tool is self-cancellation, and that only integrating into specific business assets for monetization can make commercialization work. This is a personal perspective with reference value for assessing AI4S investment timing.

10. macOS Local Knowledge Graph Contains 181 Entity Types, Including Family Relationship Profiles

Your Mac has been building a file on you all along, including family relationship scores. Eighth Rank researchers reverse-engineered macOS and found 51 SQLite databases locally , with the main graph using 181 entity types and 278 predicates to record provenance, confidence, and timestamps. Two sets of XGBoost classifiers use 144 features to infer family relationships, with scores stored separately in entityTagging.db across 17 relationship columns. The databases are stored in plaintext, with the directory protected by TCC full-disk access; the research is based on macOS 26.6.2, and the relevant components have existed since macOS 13 🔒.

⌘ Top Open Source Projects

dsx-ai-factory/nvke: Kubernetes Management Plane for Bare-Metal GPU Clusters

A common pain point when adding bare-metal GPU nodes to a Kubernetes cluster is the lack of a unified scheduling layer. dsx-ai-factory/nvke publishes the core code for a Kubernetes management plane for NVIDIA GPU bare-metal nodes , written in Go. It gained 345 Stars today, bringing the total to 851 Stars. Worth trying for MLOps engineers who run their own GPU clusters and are evaluating bare-metal K8s solutions.

anthropics/claude-quickstarts: Quick-Start Deployable Apps for the Claude API

For developers who want to build a real running application with the Claude API, anthropics/claude-quickstarts provides a collection of Python examples ranging from API integration to deployable applications . It gained 32 Stars today, bringing the total to 19,416 Stars. A good starting point for developers new to the Claude API or engineers looking to quickly validate an integration approach.

google-deepmind/mujoco: DeepMind’s General-Purpose Physics Simulator

When training robots or reinforcement learning agents, the accuracy of the physics simulation environment directly affects Sim-to-Real transfer performance. google-deepmind/mujoco is a multi-joint contact dynamics simulator , implemented in C++, gaining 41 Stars today for a total of 16,418 Stars. Suitable as a simulation foundation for teams working on robot control, motion planning, and embodied intelligence research.

◉ Social Media Picks

Google DeepMind Teases Major AI for Science Announcement Coming Within Weeks

DeepMind VP hints at a phase-transition-level breakthrough about to debut. DeepMind VP of Research Pushmeet Kohli posted that several of his team’s projects are experiencing a “real phase transition” in the speed of scientific discovery; Gorden Sun relayed this preview of an upcoming DeepMind AI for Science announcement on X, noting that publications “in the coming weeks” will offer a glimpse. This is a personal preview, not an official product announcement; the specific direction and release timeline have not been disclosed. Researchers following AI for Science can mark this time window.

😄 AI Fun Facts

Annoyed by Too Many Changes

Caleb has been writing code with an AI coding assistant for a long time and counted that he released 244 versions over the course of a year, with more than three-quarters of them involving changes to system prompts and tool descriptions. He switched to Pi, wrote a cross-tab text-selection extension in about 20 minutes, and documented the process of migrating from Claude Code to Pi . Annoyed by others making too many changes — turns out he was even quicker to make them himself.

❓ Related Questions

When Gemini Is Running a Coding Task, How Do You Verify It Actually Understood the Requirements Rather Than Pattern-Matching?

Change one small detail and have it solve the problem again. MaxxBench’s requirement-perturbation test found that Gemini Flash scores dropped sharply, showing that small changes can expose whether a model is regurgitating answers. To try it yourself, you can start testing with 大爱·AI账号店’s Gemini Pro ready-to-use accounts, 3–12 months, randomly assigned — these are independent ready-to-use accounts, not renewals of old ones.