仓库/Skill 库

学习驱动运动预测与规划中风险校准及运行时保障系统综述的可复现材料。Reproducibility materials for a systematic review of risk calibration and runtime assurance in learning-enabled motion prediction and planning.

AI DeFi Agent 的执行前守卫 API —— <300ms 返回二元判定，通过 x402 支付 0.01 USDC。Pre-execution guard API for AI DeFi agents — binary verdict in <300ms, $0.01 USDC via x402

某私有组织的公开主页，包含多个 .git 大学项目，涵盖软件工程、全栈 Web 开发、数据科学与工程（AI/ML）、嵌入式系统、GIS 等先进计算技术，分别在 Uppsala、Canterbury、Auckland 完成，并含多次海外正式交换经历；多数项目涉及广泛协作，且均以 distinction 等级通过Public profile for a private org containing multiple .git university projects in software engineering, full-stack web dev, data science & eng (AI/ML), embedded systems, GIS. & other advanced computing tech completed at Uppsala, Canterbury, Auckland & formal exchanges abroad; many involving extensive collaboration, all averaging with distinction

中文 MCP Apps 静态就绪与安全检查 CLI

🔐 使用 KeyVault Pro 安全管理 API key，具备军事级加密、智能自动填充与面向开发者的无缝 AI 集成。🔐 Manage API keys securely with KeyVault Pro, featuring military-grade encryption, intelligent auto-fill, and seamless AI integration for developers.

一个为 Model Context Protocol 提供策略执行能力的 gateway。聚合 MCP server，强制工具调用策略，扫描结果中的 prompt injection 与凭证泄露，并审计一切。Rust 实现。A policy-enforcing gateway for the Model Context Protocol. Aggregates MCP servers, enforces tool-call policy, scans results for prompt injection and leaked credentials, and audits everything. Rust.

在小型 LLM 中植入认知美德的激活引导实验——三项发现与一个失败模式数据集（DOI 见 README）Activation-steering experiments on installing epistemic virtues in small LLMs — three findings + a failure-mode dataset (DOIs in README).

使用 TurboQuant 和 MLX 在 Windows 上运行 Qwen3.5 语言模型，实现快速的本地推理。Run the Qwen3.5 language model on Windows using TurboQuant and MLX for fast local performance.

2026 年 8 大 AI 编码 Agent 防护面盘点 — Sigil Hook 与 Sandbox 评测8 Best AI Coding Agent Guard Surfaces 2026 – Sigil Hook & Sandbox Review

中立的 MCP server 信任注册中心——连接前先查验。Neutral trust registry for MCP servers — check before you connect

SURF2026 计算金融项目，研究感知自由边界的美式期权风险曲面，使用 CN/PSOR 基准、文献综述、实验报告以及 Codex 辅助的分步研究规划SURF2026 computational finance project on free-boundary-aware American option risk surfaces, using CN/PSOR benchmarks, literature review, experiment reports, and step-by-step Codex-assisted research planning.

无凭据 MCP 策略代理，具备细粒度能力与防篡改凭证。Secretless MCP policy proxy with scoped capabilities and tamper-evident receipts

本项目研究金融市场的统计特性、不同资产的收益行为、风险测度以及跨市场关联的量化方法，奠定量化研究的核心技能基础。This project investigates the statistical properties of financial markets, how different assets behave, how risk is measured, and how inter-market relationships can be quantified; building the core skills used in quantitative research.

提供清晰的分步指南，介绍如何在 Arch Linux 上结合 rEFInd 与 Nvidia 硬件启用带模块签名强制的 Secure Boot。Provide a clear, step-by-step guide to enable Secure Boot with module signature enforcement on Arch Linux using rEFInd and Nvidia hardware.

面向编程 Agent 的确定性业务变更安全闸门——基于证据的 BLOCK/ALLOW 决策。Deterministic business-change safety gate for coding agents — evidence-backed BLOCK/ALLOW decisions

受治理的持久化 AI worker，配备默认拒绝权限、人工审批、dry-run 以及经过认证的 MCP。Governed persistent AI workers with default-deny authority, human approvals, dry-runs, and an authenticated MCP.

基于 Base 的 fail-closed 支付安全判定：识别已知恶意钱包、高度相似 token 及反复 rug 的资方。非托管、仅依赖描述符——工具本身不转移资金。Fail-closed safe-to-pay verdicts on Base: known-bad wallets, look-alike tokens, repeat rug funders. Non-custodial, descriptor-only — no tool moves funds.

插入到 MCP 服务器前方的安全与可信代理——策略拦截、rug-pull 检测、JIT 审批、哈希链审计。124 个测试。MCP 서버 앞에 끼어드는 보안·신뢰성 프록시 — 정책 차단, rug-pull 탐지, JIT 승인, 해시 체인 감사. 124 tests

基于 TypeScript 的独立框架，通过统一原语 Component 组合安全的个人 AI agent：manifest-first 发现、能力授予、正交模型路由与信息流控制。Standalone TypeScript framework for composing secure personal AI agents from one universal primitive — the Component. Manifest-first discovery, capability grants, orthogonal model routing, and information-flow control.

针对 PX4 飞控软件的覆盖率引导事务语义审计研究 artifactResearch artifact for coverage-guided transactional semantic auditing of PX4 flight-control software.

可验证混合 AI 执行的合约层 —— 研究预标准 0.36.0。A contract layer for verifiable hybrid AI execution — research pre-standard 0.36.0

KARZOUN-X —— 在地球通信延迟或不可用条件下，面向航天器自主故障诊断的资源感知、安全门控本地 AI 开源研究。KARZOUN-X — Open research on resource-aware, safety-gated local AI for autonomous spacecraft fault diagnosis under delayed or unavailable Earth communication.

Astro 静态站点，英意双语，安全设计（security-by-design），AI 辅助的月度 RAG 杂志。部署于 Cloudflare。Astro (static), bilingual EN/IT, security-by-design, AI-assisted monthly RAG magazine. Deployed on Cloudflare.

用于实证金融研究的 Python 流水线，集成数据获取、统计分析、尾部风险分析与投资组合构建。A Python pipeline for empirical financial research, integrating data acquisition, statistical analysis, tail risk analytics, and portfolio construction.

本仓库包含论文补充材料，旨在提升系统综述的透明度、可复现性与完整性，涵盖筛选文档、检索策略以及计算机视觉中物理对抗攻击的详细分析。This repository contains the supplementary materials accompanying our paper. The materials are designed to support the transparency, reproducibility, and comprehensiveness of the systematic review, including screening documentation, search strategies, and detailed analysis of physical adversarial attacks in computer vision.

为 AI agent 提供持久且安全约束的浏览器控制——使用语义化 CLI 替代 Playwright 脚本或屏幕坐标。通过统一接口，背后使用 macOS WKWebView 与 Linux ChromiumPersistent, safety-enforced browser control for AI agents — semantic CLI instead of Playwright scripts or screen coordinates. macOS WKWebView + Linux Chromium behind one contract.

大语言模型监督微调的反馈对齐方法Feedback alignment for supervised fine-tuning of large language models

本地 CLI，用于发现本地网络上的影子 MCP 服务器并审计认证状态。零遥测，纯本地。Local CLI to discover shadow MCP servers on your local network & audit authentication status. Zero telemetry, local-only.

Kiki — 一个面向 Google Campaign Manager 360 的开源对话式 trafficking 助手。集成 70 个 CM360 工具，具备写入安全确认、内置 QA 以及零凭证 MCP demo。Kiki — an open-source conversational trafficking assistant for Google Campaign Manager 360. 70 CM360 tools, write-safety confirmation, built-in QA, zero-credential MCP demo.

使用自然语言查询分析并比较 700+ DeFi 金库的风险评分，无需 API key 或安装Analyze and compare risk scores of 700+ DeFi vaults using natural language queries without requiring API keys or installations.

硕士论文（KCL）：在攻击者-机器留出划分下重新衡量 IoT 入侵检测，并测试基于 LLM 的报文预测是否能提升性能。移除泄露后，macro-F1 从 0.90 降至 0.60。MSc dissertation (KCL): re-measuring IoT intrusion detection under an attacker-machine holdout, and testing whether LLM-based packet prediction improves it. macro-F1 0.90 -> 0.60 once leakage is removed.

Shadow / Unmanaged API 的发现、分类、安全验证与可复现研究基准。Shadow / Unmanaged API discovery, classification, security validation, and reproducible research benchmarks

可执行真实操作的 AI agent 层级——shell、GitHub、邮件、采购——以代码编写（非配置）的审批底线（Latch）作为门槛。零依赖。A hierarchy of AI agents that take real actions — shell, GitHub, email, purchases — behind an approval floor written in code, not configuration (Latch). Zero dependencies.

可复现研究原型：面向有状态云服务的碳感知与 SLO 约束自动伸缩A reproducible research prototype for carbon-aware SLO-constrained autoscaling of stateful cloud services using synthetic workload traces, a threshold baseline, and a predictive risk-aware autoscaling policy.

应用型 AI 网络安全研究实验室，涵盖机器学习威胁检测、对抗机器学习、鲁棒性、网络威胁情报、MITRE ATT&CK/ATLAS、LLM 安全与可复现研究。Applied AI Cybersecurity research labs covering machine learning threat detection, adversarial ML, robustness, cyber threat intelligence, MITRE ATT&CK/ATLAS, LLM security and reproducible research.

面向系统综述数据抽取、证据定位与偏倚风险评估的冻结式（frozen）防泄漏基准。Frozen, leakage-aware benchmark for systematic-review data extraction, evidence localization, and risk-of-bias support