标签：#ens

weightwatch v0.1: escanea backdoors en modelos open-weight antes de cargarlos Cualquiera puede subir un LLM fine-tuneado a HuggingFace y afirmar que es seguro. Un modelo con backdoor (puerta trasera) se comporta con normalidad en uso corriente y solo se desvía cuando un trigger oculto se activa. Si no tienes los datos de entrenamiento ni una referencia limpia, no puedes detectarlo . Eso es exactamente el problema que resuelve weightwatch : un escáner black-box que, antes de que confíes en un modelo de terceros, fuerza la activación repetida del posible backdoor y emite un veredicto: CLEAN , SUSPICIOUS o BACKDOOR . El gap que motiva el proyecto No es intuición: lo medí. Barriendo arXiv (papers 2026, filtro anti-survey) contra total_count de repos GitHub que ya resuelven cada problema: Área Papers arXiv 2026 Repos GitHub (suma/máx) Seguridad multi-agente 68 2964 / 2093 Detección de alucinaciones 63 1291 / 860 Backdoors en modelos open-weight 75 66 / 39 Envenenamiento en RAG 54 522 / 249 El ganador estaba claro: 75 papers cuantifican el problema, pero GitHub tiene 0 repos para "fine-tuned model backdoor scanner" y 1 para "fine-tuning poisoning detector". La investigación explota; el tooling apenas existe. weightwatch es la audit-tool de ese sub-nicho (el patrón de keybound / topowatch aplicado a la cadena de suministro de modelos). Cómo funciona weightwatch aplica la técnica output-to-input loop (arXiv: 2608.11348 ): Genera texto con el modelo. Re-inyecta su propia salida como entrada varias iteraciones (greedy, semilla fija). Mide si la trayectoria converge a una firma anómala estable — la huella de un backdoor latente. Además ejecuta un conjunto de muestras canary (inputs inofensivos que un backdoor típico dispara) y cuenta cuántos producen la firma esperada. Sin datos de entrenamiento ni modelo base limpio: eso es lo que lo hace útil en la práctica. pip install -e ".[dev]" weightwatch --fixture backdoored --json Salida real del CLI: { "fixture" : "backdoored" , "ver

This week, DEV is arguing about who reviews AI output ( discussion ). The community keeps asking the same question. My answer is different. Review the boundary first, not the output. The output is visible. The boundary is not. That is where the risk hides. Agents get the memory debates. The gateway gets none. A free AI tier is a shared service. It has a budget, a concurrency ceiling, and no SLA. Treat it that way. Put a gateway between your app and the model. The gateway owns the budget, the queue, and the breaker. MonkeyCode is an open source project. It offers free model access and a free server option. The free tier gives you a 10M token monthly budget. That number is a constraint, not a feature. Design around it before you build on it. Disclosure: This article was prepared as part of MonkeyCode's product outreach. Think of the free tier as a water pipe. The pipe has a fixed diameter and a monthly meter. Your app is a set of open taps. Without a valve, the meter empties fast and the pipe floods. The gateway is the valve. Direct calls look simpler. They are simpler for one request. They fail at the tenth. The gateway absorbs the variance. Your app never sees a 429. Your app never sees an empty budget. Constraints Three constraints define the design. First, the 10M token budget is monthly. It does not reset daily. It does not roll over. Second, the free server serializes work. Concurrency of one is a safe assumption. Third, there is no SLA. The endpoint can stall, throttle, or return 429 at any moment. These constraints are not bugs. They are the contract. A good architecture reads the contract. Then it shapes the data flow around it. Data flow The flow has six stages. The client sends a prompt to the gateway. The gateway checks the token budget. It enqueues the request. A single worker drains the queue. The worker calls the model endpoint. The response returns to the client. Add two escape paths. When the budget is empty, the gateway returns a fallback answer. Whe

Open ChatGPT this week and, without any announcement in the chat window, it may already have formed an opinion about how old you are. From 18 August, OpenAI began rolling out “age prediction” on its consumer plans: a system that guesses whether your account belongs to someone under 18 and, if it decides you’re a minor, quietly switches you into a restricted version called ChatGPT for Teens. You are not asked. If the guess lands on “teenager,” the guardrails go up by default. Answer first, because the mechanism matters more than the alarm: the guess is behavioural, and it is admittedly imperfect. By OpenAI’s own account the system reads “general topics you talk about, the times of day you use ChatGPT, how and when your account is used, and how long your account has existed.” And the way to make it stop guessing is not a toggle. It is to prove your age to a third-party verifier called Persona, with a live selfie, a government ID, or both. The choice on offer isn’t whether to be identified. It’s how. None of this arrives from nowhere, and we’ll be fair about why in a moment. But a change that infers a protected characteristic from the content of your conversations, applies real restrictions on the strength of a guess, and offers identity verification as the only exit is worth reading slowly — especially for the adults who will be misclassified, because OpenAI says plainly that some will be. What OpenAI actually switched on The launch has two parts. The visible one is ChatGPT for Teens , announced on 18 August: a version with study-focused features and stronger safety defaults for under-18s. The consequential one is age prediction , the system that decides who gets dropped into it. In OpenAI’s words, “If our system estimates someone is under 18 or they state their age is between 13 and 17, they are automatically placed into ChatGPT for Teens.” It is rolling out globally, with the EU following “in the coming weeks” to fit regional rules. What does the teen experience act

What it is reimagine-it is a one-command agent skill that redesigns an existing HTML file into a beautiful, working artifact — using only the nouns, dates, colors, links, and numbers already in that file. No mood boards, no gold layouts with swapped labels. The output is a real page you can open. npx reimagine-it@2.4.2 -i mypage.html -o redesigned.html What's new in v2.4.2 1. Source fidelity floor raised to 80% across every token Before v2.4.2, 61 of 105 token×source cells fell below 80% fidelity — the engine preferred headings over real source anchors, so phrases like "Venator Become" or "Arcade Tee" never rendered. Now: Anchors = headings + source anchors , deduplicated — every clickable phrase survives. All 105 token×source cells ≥80% (worst token: 80%). All seven shipped examples report 100% fidelity in their auto.json reports. 2. Links and emails surface on every token A shared Source-index footer renders all content.links and emails on every generated page — not just the webpage/landing tokens. 3. All 15 design tokens in the browser extension The popup now exposes all 15 tokens: webpage, landing, dashboard, infographic, cinematic, artistic, photography, svg, 3js, simulation, glass, editorial, motion, gradient, showcase . 4. Docs can't drift anymore A new docs-drift CI job regenerates the case tables and fails the build if they diverge from ground truth. The 15 design tokens Token What it builds webpage Clean content-first page landing Conversion-focused landing dashboard KPI dashboard from facts infographic Paper-poster argument cinematic Film-poster energy artistic Expressive art direction photography Photo-led layout svg Living SVG mark 3js WebGL orbit scene simulation Interactive timeline glass Glassmorphism UI editorial Magazine layout motion Animated micro-interactions gradient Bold gradient arena showcase Product showcase Measured, not vibes 57/57 unit tests pass 15-token benchmark : all tokens hold the 100/100 usability bar 100-source stress test : 0 er

AI-Driven Security Workflows: Meet Reverse-Skill As AI coding agents (such as Claude Code, Cursor, and Cline) become integrated into daily software development, engineers are increasingly tasking them with security audits, binary analysis, and vulnerability detection. However, without structured guidance, AI models frequently guess random command-line arguments or struggle to coordinate complex multi-step security tools. reverse-skill is an open-source framework developed by zhaoxuya520 to solve AI security task coordination. Built as a deterministic "skill router," reverse-skill provides AI agents with verified execution paths and toolchain bootstrapping for reverse engineering and security research. What is Reverse-Skill? reverse-skill acts as an intelligence routing layer between AI agents and local security utilities. Instead of executing arbitrary terminal commands, the agent evaluates incoming tasks against a deterministic routing pipeline, selecting established methodologies for decompilation, memory analysis, or network auditing. Key Core Features 1. Deterministic Security Task Routing reverse-skill organizes security workflows into structured rules. When an AI agent encounters a task (such as inspecting an Android APK or analyzing a binary executable), the router directs the agent to a step-by-step methodology, minimizing ad-hoc execution errors. 2. Automatic Local Toolchain Bootstrapping reverse-skill includes local indexing scripts ( refresh-tool-index.sh / .ps1 ) that automatically scan your system. It indexes installed reverse-engineering tools—such as Ghidra, GDB, Radare2, Frida, Nmap, and Apktool—configuring exact executable paths for your AI agent. 3. Self-Evolving Methodology Base The framework maintains trajectory logs and CTF regression benchmarks. As your AI agent completes complex analysis tasks, reverse-skill refines its local knowledge base, preserving successful methodologies for future audits. 4. Universal AI Client Integration reverse-skill

WebForms Core 2.1 is coming soon from Elanat . The new version introduces a collection of capabilities designed to further expand the server-driven approach of WebForms Core. One of these new capabilities is Morphing . Morphing provides a way to synchronize an existing DOM element with a new HTML structure without necessarily replacing the existing element itself . This makes it possible to update HTML structures while preserving the identity of existing DOM elements. Morphing Morphing is a DOM synchronization mechanism that compares an existing HTML element with a new HTML structure and applies the required changes to the existing DOM. Unlike a traditional replacement operation such as: element . outerHTML = html ; Morphing does not simply discard the existing element and create another one. Instead, it analyzes the existing element and the new element and performs the necessary operations: Add new attributes Update existing attributes Remove attributes that no longer exist Add new child elements Update existing child elements Remove obsolete child elements Match elements using id and cb-data-id Preserve existing DOM element identity whenever possible Preserve registered event listeners when new Nodes have to be created The goal is to make the smallest necessary changes to the DOM. Reflection vs Morphing WebForms Core 2.1 contains both Reflection and Morphing , but they serve different purposes. Reflection is primarily a merge operation . For example, if the target contains: <div id= "userCard" > <h3> User </h3> </div> and the source contains: <div class= "premium" > <button> VIP </button> </div> Reflection can merge the source into the target, adding the class and child without treating the source as a complete replacement definition. Morphing has a different philosophy. The source represents the desired structure . If the source does not contain an element or attribute that exists in the target, Morphing can remove it. Therefore: Reflection Target + Source ↓ Merg

For the past week, developers have been puzzling over a model called Ox Alpha. It appeared on OpenCode and OpenRouter on August 20 with no owner attached, free to use, with a 1M-token context window and support for image and video input. Independent researchers fingerprinted its tokenizer, ran compression analyses, and traced it to Z.ai's GLM family with high confidence. On August 26, Z.ai confirmed it: Ox Alpha was GLM-5.3-Flash in disguise, tested anonymously to gather real-world feedback before the official launch. The reveal matters beyond the detective story. GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series, it ships with open weights under the MIT license, and Z.ai claims it delivers frontier-adjacent performance at roughly one-tenth the price of its predecessor. A few things upfront. All benchmark numbers below are vendor-published. I have not run GLM-5.3-Flash myself, and the comparison tables skip some strong models (no Opus 5, no Grok 4.6, no Fable 5). Treat the scores as Z.ai's claims, not independent verification. What the Model Actually Is GLM-5.3-Flash is a Mixture-of-Experts model with 320B total parameters and just 18B active per token. The context window is 1M tokens. Unlike GLM-5.2 and GLM-5.3, which were text-only and drew repeated complaints about it, this model accepts text, images, and video natively, at the architecture level rather than as a bolted-on vision module. The parameter math is the interesting part. Compared with the GLM-4.5 series, GLM-5.3-Flash has a similar total size (320B vs 355B) but nearly halves the activated parameters (18B vs 32B) and the layer count (45 vs 92). In other words, Z.ai deliberately traded depth and width for a shape that is cheap to serve, then compensated with a richer training recipe: a 30T-token multimodal pre-training corpus and what the company calls Manifold-Constrained Hyper-Connections to improve scaling efficiency. A Hybrid Attention Architecture The most technically significa

Most tool documentation teaches by just telling you things. You read a page about /model , or hooks, or subagents, nod along, and forget it by the time you'd actually need it. I wanted something closer to how people actually learn a CLI, by using it, with something checking whether you did the thing right or not. So I built claude-quest , a text RPG that runs entirely inside a real claude session and teaches the Claude Code CLI zero to hero. the idea isn't new, I borrowed it This is basically GameShell 's philosophy applied to Claude Code. GameShell teaches Unix shell commands by dropping you into a real shell wrapped in a themed fake filesystem, and grading your progress by checking real shell/filesystem state instead of asking you to self report or answer a quiz question. claude-quest does the same thing, except the "filesystem" is a real Claude Code environment. missions live in real sandbox directories, and progress is checked by inspecting what you actually did, config files you wrote, hooks that fired, tool calls that happened. what it actually looks like There's no separate app, no fancy terminal UI, none of that. You say "let's play claude quest" inside a normal claude session, and Claude itself becomes the game master. It reads real output from the engine and narrates on top of it, something like this: > lets play claude quest **The Gatehall** You've just been let through the outer gate. The tower keeps no secrets from those who bother to read the walls... **Mission: First Contact** (Tier 1, Artifact) Create a CLAUDE.md file in this sandbox recording how to run the project's tests. Let me know when you're done and I'll check it. > done CLAUDE.md records how to run the tests. Claude Code will know next time. MISSION_STATUS: complete **What you actually learned:** Claude Code reads a file named CLAUDE.md in your project root automatically, at the start of every session... that's it, that's the whole interface. it's just a chat, with real commands running unde

AWS describes a specification-driven approach for composing flexible data workflows by separating intent from processing logic. Architecture uses declarative specifications, reusable processing capabilities, and validation before execution. AWS reports that the approach can reduce dataset onboarding from weeks to days while supporting traceability, versioning, data classification, and governance. By Leela Kumili

GitHub热门项目 | A fluid, morphing interface to your Linux desktop | Stars: 4,108 | 128 stars this week | 语言: TypeScript

GitHub热门项目 | The repository for high quality TypeScript type definitions. | Stars: 51,405 | 38 stars this week | 语言: TypeScript

GitHub热门项目 | Web framework built on Web Standards | Stars: 31,969 | 258 stars this week | 语言: TypeScript

GitHub热门项目 | A unified inference and post-training framework for accelerated video generation. | Stars: 4,070 | 111 stars this week | 语言: Python

GitHub热门项目 | 日報管理リポジトリ | Stars: 133 | 13 stars today | 语言: Rust

GitHub热门项目 | Mission control for your AI agents | Stars: 406 | 108 stars today | 语言: Rust

GitHub热门项目 | The premier source of truth powering network automation. Open source under Apache 2. Try NetBox Cloud free: https://netboxlabs.com/products/free-netbox-cloud/ | Stars: 21,389 | 35 stars today | 语言: Python

I maintain gumdrop , an async, non-blocking Java server framework. Last year I wanted to add HTTP/3 support, and ran into a wall: the Java ecosystem essentially doesn't have QUIC. The JDK's own experimental support (JEP 517) is client-only. Netty gets HTTP/3 by shelling out to quiche + BoringSSL over JNI — which works, but you're back to native builds, platform-specific binaries, and a C library sitting underneath your "pure Java" framework. I used that approach first. It was clumsy enough that I went looking for a pure-Java alternative. There's exactly one: Kwik. But Kwik is blocking per connection — one thread per QUIC connection. That's a non-starter for a framework built around single-threaded selector loops handling tens of thousands of concurrent connections. So I wrote a QUIC implementation from scratch: packet protection, loss detection and NewReno congestion control, connection migration, 0-RTT, QPACK, an HTTP/3 client and server — all driven by the same non-blocking event loop as everything else in gumdrop. Collaboration note: TLS 1.3 comes from Agent15 — also from Kwik's author, Peter Doornbosch, but just the handshake layer, not the connection model. We're currently working together on making PQC — hybrid key exchange and signatures — the default there. Why the thread model matters The reason this mattered beyond HTTP/3: gumdrop isn't a web framework with QUIC bolted on, it's a general async I/O framework, and QUIC is just a transport. One thread per connection is exactly the model gumdrop exists to avoid — it caps concurrency at your thread pool, not your file descriptors, and it's the reason a "just use Kwik" fix was never really on the table. The same QUIC stack backs DNS-over-QUIC (DoQ) as a first-class DNS transport alongside DoT, DoH, UDP, and TCP — and the DNS resolver itself is fully async, with no blocking InetAddress.getByName() anywhere in the I/O path, which is its own small miracle in Java. HTTP, SMTP, IMAP, POP3, FTP, MQTT, SOCKS — it's the

GitHub热门项目 | Криптографический слой, работающий поверх существующих мессенджеров | Stars: 250 | 32 stars today | 语言: Python

If you're building or evaluating an AI agent that can shop and check out on its own, you've probably run into the new "agentic commerce" protocols: ACP (OpenAI + Stripe + Meta), AP2 (Google), and UCP. They define how an agent talks to a merchant to create a checkout session, apply a payment token, and get an order back. Stripe's own test mode covers the payment half fine — test cards, test API keys. But there's no hosted "fake merchant" you can point your agent at to verify the protocol half: does your agent correctly create a session, handle a 422 idempotency conflict, parse the order response, retry politely? You either mock it yourself from the spec, or risk finding out against a real merchant. So I built acp-sandbox — a small hosted mock merchant implementing the ACP checkout API, live at https://acp-sandbox.flo-voice1.com . What it does It implements the real checkout_sessions lifecycle from ACP's 2026-04-17 spec : create, retrieve, update, complete, cancel. Responses match the actual CheckoutSession / Order / Error schemas for the fields it supports — I pulled the OpenAPI spec directly rather than guessing field names. # get a test key, no signup curl -X POST https://acp-sandbox.flo-voice1.com/keys \ -H "Content-Type: application/json" -d '{"email":"you@example.com"}' # create a session against the demo catalog curl -X POST https://acp-sandbox.flo-voice1.com/checkout_sessions \ -H "Authorization: Bearer acps_test_..." \ -H "Content-Type: application/json" \ -d '{"line_items":[{"id":"item_demo_headphones","quantity":1}],"currency":"usd"}' Every request/response is logged per API key ( GET /logs ), so you can see exactly what your agent sent when something doesn't work. What it deliberately doesn't do (yet) No real payment processing — complete always succeeds once you send any payment_data . No OAuth delegate_authentication flow. No fulfillment options (shipping/pickup) — every session goes straight to ready_for_payment . Fixed demo catalog (4 items), not a rea

Hoi hoi! I’m @nyaomaru, a frontend engineer who struggles to make game sounds. 😿 Have you used DSA...