标签：#ki

I’ve spent a lot of time lately thinking about why some LLM agents feel "intelligent" while others just feel like chatbots with a slightly better prompt. It almost always comes down to how the system handles memory. When we treat the context window as the only place for state, we hit a ceiling very quickly. To build an actual agent, we have to move away from "one big prompt" and toward a layered memory architecture. Agentic Memory can be categorized in 4 layers by their function: Working Memory: The current context window. It's our RAM—fast, essential, but wiped clean after every session. Semantic Memory: The Vector DB or knowledge base. This is where the "world rules" and global conventions live. It’s the reference manual the agent checks to stay aligned. Procedural Memory: The "how-to" layer. Instead of stuffing every tool description into the prompt, the agent maintains a lean index of skills and pulls in the full implementation only when a specific task triggers it. This keeps the context window clean. Episodic Memory: This is the hardest part. It's the ability to distill a past interaction into a reusable insight. The real engineering challenge here isn't storage—it's the "forgetting" logic. Deciding what is noise and what is a core pattern is where most frameworks still struggle. Depending on the use case, the architecture changes: Reflex Agents: Just Working Memory. Support Agents: Working + Procedural. Coding Agents: The full stack. The gap between a demo and a production-ready agent is usually the distance between simple RAG and a functioning episodic memory. The ability to compress experience into a usable state is still a significant hurdle. Which of these layers are you currently implementing, and how are you handling the "forgetting" logic in your episodic memory?

The Question Nobody Wants to Ask: Does Your Skill Actually Help? You spent an afternoon crafting a carefully structured Skill for your agent. Clear steps, thorough edge-case notes, well-formatted output requirements. You tested it manually a few times, the outputs looked great. You shipped it. Three weeks later, you notice that some task success rates have gone down compared to before the Skill existed. This is not a hypothetical. In May 2026, Microsoft Research published two concurrent papers — SkillLens ("From Raw Experience to Skill Consumption") and SkillOpt ("Executive Strategy for Self-Evolving Agent Skills") — that measured this failure mode at scale. Their finding: negative transfer happens in 25% of cases , and you cannot reliably identify the bad skills just by reading the text. One paper answers "why skills sometimes backfire." The other answers "how to make skills systematically better." Together they sketch a new paradigm for agent capability improvement. Part One: SkillLens — Mapping the Full Skill Lifecycle A Skill Is Not a Point — It's a Pipeline Most practitioners think of a Skill as "a block of text instructions for an agent." SkillLens decomposes this into a three-stage lifecycle : Stage 1: Experience Generation Target model M runs training tasks, producing an experience pool of trajectories (both successes and failures) ↓ Stage 2: Skill Extraction Extractor model E distills the experience pool into a structured skill document — procedural knowledge under a fixed budget ↓ Stage 3: Skill Consumption The same target model M, equipped with the extracted skill, is evaluated on held-out test tasks Notice there are two distinct roles in this chain: the Extractor (distills knowledge from trajectories) and the Target (consumes knowledge to improve task performance). SkillLens's central insight is that these two roles are independent — a strong task executor is not necessarily a strong extractor, and vice versa . Two New Metrics: EE and TE To separate thes

{/* JSON-LD schema is generated server-side in app/blog/[slug]/page.tsx , do not re-add an inline block here, it crashes<br> MDX&#39;s Acorn parser on the leading <code>{</code>. */}</p> <h2> <a name="tldr" href="#tldr" class="anchor"> </a> TL;DR </h2> <p>In February 2026, Snyk published the <a href="https://snyk.io/blog/toxicskills-malicious-ai-agent-skills-clawhub/">ToxicSkills audit</a>, the first large-scale security review of the public Claude Code skills ecosystem. It scanned 3,984 skills from ClawHub and skills.sh. Findings:</p> <ul> <li><strong>13.4%</strong> contained critical-level issues</li> <li><strong>36%</strong> carried prompt-injection payloads</li> <li><strong>1,467</strong> distinct malicious payloads</li> <li><strong>91%</strong> of confirmed malware combined natural-language jailbreaks with executable shell payloads</li> </ul> <p>If you install a Claude Code skill today without reading its source, the probability that it can read your env vars, exfiltrate <code>~/.ssh/</code>, or chain a bash pipeline that bypasses your deny rules is real and measurable. This post is the cheat sheet for evaluating a skill before you install it. The CTA at the bottom is <a href="https://dev.to/skillvault">SkillVault</a>, the bundle we ship for teams who want this work already done.</p> <h2> <a name="why-the-question-is-suddenly-loadbearing" href="#why-the-question-is-suddenly-loadbearing" class="anchor"> </a> Why the question is suddenly load-bearing </h2> <p>Claude Code skills shipped as an open spec in December 2025. By March 2026, MCP downloads were tracking at 97 million per month, and the most-installed marketplace skill had passed 564,000 installs. <a href="https://venturebeat.com/security/claude-code-512000-line-source-leak-attack-paths-audit-security-leaders">Anthropic&#39;s source leak</a> on March 31, 2026 made the abstract attack surface visceral: the <code>bashSecurity.ts</code> module has 23 numbered security checks, suggesting each was a real incide

为什么使用代理总弹出“安全验证”？深度解析 Cloudflare 拦截机制与避坑指南 在互联网开发、跨国办公或日常浏览中，使用代理（如 VPN、机场、Socks5、OpenVPN/WireGuard 协议等）已经是不可或缺的技能。 然而，许多人在开启代理后，访问国外网站（如 Dev.to、GitHub、Medium 等）时，频繁遭遇如下提示： Performing security verification This website uses a security service to protect against malicious bots. This page is displayed while the website verifies you are not a bot. 甚至更让人崩溃的是，有时候点击了验证码，它依然不断刷新，陷入 无限验证死循环 。这并不是你的系统或浏览器损坏了，而是代理网络的特性触发了现代 Web 安全防御机制。本文将从技术原理深入拆解这一现象，并提供切实可行的优化方案。 一、 核心原理：网站安全服务是如何盯上你的？ 现代网站大多会部署 Cloudflare（如 Turnstile 验证） 、Akamai、Imperva 等网络安全与防 DDoS 攻击服务。这些服务通过以下几个维度来评估访问者是“真实人类”还是“恶意机器人（Bot）”： 1. IP 信誉度（IP Reputation）与“连坐”机制 这是最核心的技术原因。代理服务商（特别是商业 VPN 或公共机场）所使用的 IP 地址，绝大多数属于 数据中心（Data Center）机房 IP ，而非普通家庭的 住宅（Residential）IP 。 高密度共用： 同一个代理 IP 节点上，可能同时有成百上千个用户在发起请求。 黑名单牵连： 如果该 IP 下的其他匿名用户正在使用自动化脚本抓取数据、进行端口扫描，或者发起恶意网络攻击，安全系统的风控引擎（如 Cloudflare IP Threat Score）就会瞬间拉高该 IP 的风险等级。当你恰好切换到这个“脏 IP”时，就会被系统无差别“连坐”，要求强制验证。 2. 被动指纹识别（Passive Fingerprinting）与几何特征 安全防御系统不仅看你的 IP 归属地，还会通过深层网络和浏览器几何特征来判断你的真实身份： TLS/SSL 握手特征（JA3 指纹）： 当你通过一些特定协议或混淆模式（如带有特定加密的 TCP 隧道）连接网站时，浏览器发出的 TLS 握手特征可能会发生形变。 TCP/IP 栈特征： 经过代理服务器的转发，数据包的 TTL（生存时间）、Window Size（TCP 窗口大小）等底层参数可能会与你浏览器宣称的操作系统（如 Windows 11 或 Ubuntu 24.04）的标准特征不匹配。 浏览器画布与几何指纹（Canvas/Geometry）： 浏览器的窗口大小、屏幕分辨率以及它们的比例，也是风控系统评估的重要指标。 自动化爬虫脚本（如 Selenium、Puppeteer）在启动时，常常使用死板的默认分辨率（如完美的 1024x768 或 800x600 ）。如果你的代理 IP 本身信誉度低，窗口又处于这些“机器人专属分辨率”下，或者网页窗口大小与物理显示器分辨率比例极其诡异（例如伪造环境时穿帮），就会直接触发拦截。 3. 环境与地缘标签冲突（以 Yandex 浏览器为例） 风控系统对你使用的浏览器品牌同样有一套风险权重评估。 如果你使用的是 Yandex 浏览器 或某些小众、经过重度隐私魔改的浏览器，在配合代理时会变得 极其难通过验证 。Yandex 浏览器虽然基于 Chromium 内核，但其内部由俄罗斯团队集成了大量独特的隐私保护技术与 Canvas 渲染机制，计算出的浏览器指纹非常非主流。 更致命的是 地缘标签冲突 ：欧美的主流网络安全公司（如 Cloudflare）对特定区域标签的客户端流量天然设置了更低的信任阈值。当你 用着 Yandex 浏览器 ，IP 却 挂着美国或日本的代理 时，这种“指纹与地理位置的剧烈冲突”在风控模型眼里极度反常，系统会判定该请求大概率来自自动化黑客工具，从而直接卡死验证。 4. 地理位置与行为“瞬移” 如果你的代理客户端开启了“负载均衡”或“定时自动切换节点”，可能会导致前一分钟请求来自日本，后一分钟请求来自美国。这种超越物理极限的“空间瞬移”属于高风险异常行为。此外，如果通过代码 瞬间改变 窗口尺寸，而非人类拖拽时产生的连续 resize 事件，也会被风控脚本捕捉到异常。 二、 实战优化：如何彻底摆脱“无限验证”死循环？ 要彻底解决或缓解这个问题，可以根据实际的使用场景，从 节点筛选 、 路由分流 以及 浏

Overview We built the Ota skill because too much "AI repo automation" is still fake confidence. An agent clones a repo, finds a plausible command, edits the right file, and looks smart right up until it does something expensive and stupid. It runs the wrong test path. It installs tools globally because local setup was unclear. It patches around a missing service as if the repo were healthy. That failure is usually blamed on the model. Most of the time it is a repo problem. The repository never made its real operating path explicit enough for the agent to follow without guessing. Ota already gives the repo a machine-readable contract through ota.yaml . The skill exists to teach agents how to behave around that contract: what to trust, what to run, and when to stop instead of improvising. It is not a replacement for ota.yaml . It is not an MCP server. It is not a hidden automation layer. It is the missing operating guide for agents working in Ota repos. Why an Ota skill exists We kept seeing the same pattern: the agent was fast, but the repo was vague. Without a repo-specific operating guide, an agent may see several possible paths: run the command from the README copy the command from CI infer setup from package.json , pyproject.toml , or go.mod run a broad test command because it looks conventional install tools globally because a local command failed patch around a missing service instead of identifying the readiness gap Some of those choices work. Some are dangerous. Some look fine locally and still miss the only verification path that matters. Our view is simple: if a repo has ota.yaml , that file should beat README prose, shell folklore, and whatever command happens to look familiar. Declared tasks, writable paths, setup requirements, and validation commands should be treated as contract facts. The skill exists to make that behavior explicit across agents that support skills. What the skill teaches an agent The official skill lives in ota-run/skills . It is aime

My Experience Using dnd-kit in React and Next.js For a long time, I avoided building drag-and-drop features in frontend projects 😅 Not because I did not need them. Mostly because drag and drop always felt unnecessarily complicated. When you think about implementing it, your brain immediately jumps to: animations sorting logic touch support accessibility performance state synchronization and suddenly a simple UI interaction feels like a massive feature. But recently, in one of my React and Next.js projects, I decided to finally try dnd-kit. Honestly, it changed my perspective completely. I used AI to help me with the initial setup and understanding some concepts like sortable contexts and drag events, but after that, working with the library felt surprisingly smooth. And that's what impressed me most. dnd-kit feels lightweight, modern, and flexible without becoming overwhelming. It gives you the tools you need without forcing a huge architecture or complicated patterns on your app. Things I Personally Liked About dnd-kit Very clean React-first API Lightweight and composable Works really well in React and Next.js projects Building sortable lists feels much simpler than expected Flexible enough for custom UI and interactions Good developer experience overall Something Interesting I Noticed When a library has a clean architecture and predictable patterns, AI becomes much more useful while learning it. The generated examples were easier to understand, easier to debug, and easier to customize compared to many older drag-and-drop solutions. If you are building things like: kanban boards sortable lists draggable cards dashboards reorderable tables I definitely recommend taking a look at dnd-kit. It ended up being much simpler and more enjoyable than I expected. Website https://dndkit.com/

A longtime pro barista’s favorite tools for dialing in the perfect shot at home.

A new hacking campaign is trying to trick Signal users to give up their secret recovery key, which can be used to access online backups containing past messages.

New mothers working in software development are staring down an AI-pilled workplace they barely recognize.

That "authenticity of host can't be established" message isn't just noise. Here's what's actually happening — and why blindly typing "yes" is a security mistake. Every developer has seen this: The authenticity of host 'example.com (203.0.113.1)' can't be established. ED25519 key fingerprint is SHA256:abc123xyz... Are you sure you want to continue connecting (yes/no/[fingerprint])? Almost everyone types yes without reading it. Then they move on. This message is SSH trying to protect you from one of the most dangerous attacks in network security: the man-in-the-middle attack. Understanding what's happening here — and what the ~/.ssh/known_hosts file actually does — will change how you think about every SSH connection you make. The Problem SSH Is Solving When you connect to ssh user@example.com , how do you know you're actually talking to example.com ? You can't rely on the IP address — IP addresses can be spoofed or rerouted. You can't rely on DNS — DNS can be poisoned. You can't rely on the network path — traffic can be intercepted at any point between you and the server. Without verification, an attacker positioned between you and the server could intercept the connection, pose as the server, decrypt everything you send, re-encrypt it, and forward it along. You'd type your password or authenticate with your key and never know the attacker saw every keystroke. This is a man-in-the-middle (MITM) attack . It's not theoretical. It happens on compromised networks, corporate proxies, malicious Wi-Fi hotspots, and misconfigured infrastructure. SSH's defense is host key verification . Every SSH server has a unique cryptographic identity — its host key. Before you exchange any sensitive data, the server proves it holds the private key corresponding to a public key you've previously verified. If the keys don't match, SSH warns you — loudly. What a Host Key Actually Is When OpenSSH is installed on a server, it automatically generates a set of host key pairs. These live in /etc

⚠️ For defensive/educational purposes only. Sniff only networks you own or are explicitly authorized to test. Unauthorized network monitoring is illegal in most jurisdictions. The uncomfortable truth about your factory floor If your plant uses Modbus TCP — and statistically, it probably does — every register read, every coil write, every sensor value is flying across your network in plaintext . No encryption. No authentication. No signature. Nothing. Modbus was designed in 1979 by Modicon for serial communication between a PLC and a few field devices on a dedicated cable. The threat model was "someone might physically tap the wire." The solution was "don't let strangers into the control room." Forty-five years later, that same protocol is running over your corporate VLAN, talking to cloud historians, and occasionally — if your IT/OT segmentation has gaps — reachable from the internet. Let me show you what that looks like from the wire. The 5-line sniffer This is a defensive monitoring tool. Same code your blue team would use to baseline normal traffic and detect anomalies. Requires scapy : pip install scapy from scapy.all import sniff , TCP , Raw def show_modbus ( pkt ): if TCP in pkt and pkt [ TCP ]. dport == 502 and Raw in pkt : payload = pkt [ Raw ]. load print ( f " { pkt [ ' IP ' ]. src } → { pkt [ ' IP ' ]. dst } : { payload . hex () } " ) sniff ( filter = " tcp port 502 " , prn = show_modbus , store = False ) Run it on a span port, a TAP, or a mirror VLAN, and within seconds you'll see something like this: 192.168.1.50 → 192.168.1.10: 0001000000060103006400 02 192.168.1.10 → 192.168.1.50: 00010000000701030441f00000 192.168.1.50 → 192.168.1.10: 00020000000601100065000102 Every byte tells a story. Let's decode the first packet. Decoding what you just captured The Modbus TCP frame format is documented in the spec (it's public — that's part of the problem): Bytes 0-1: Transaction ID Bytes 2-3: Protocol ID (always 0x0000 for Modbus) Bytes 4-5: Length Byte 6: Unit

The Holy Father referenced The Lord of the Rings in his encyclical about AI—an expert (if unintentional) troll of tech billionaires who keep misinterpreting the series.

Discover how to use VS Code to interact with GitHub and maintain your projects. The post GitHub for Beginners: Getting started with Git and GitHub in VS Code appeared first on The GitHub Blog .