vibehacker
News
zn Blog ·

zn opens the weights of Galvanize 3, a 140M CPU-friendly prompt-injection detector, under Apache 2.0

Published Oct 31 as usezn/Galvanize-3 on Hugging Face, it's the same encoder classifier that has screened every zn API call since Oct 23: up to 512 tokens in, one injection score out (flagged at 0.9802 or above), and it runs on CPU, so you can put it in front of your agent's tool inputs locally. zn says it scored the highest AUROC (0.965) of ten detectors on public prompt-injection benchmarks, but that comparison is its own.

Read at usezn.com

More news

View all

Etraveli opens Flight MCP, letting Claude, ChatGPT or any MCP agent search, price and book flights

Announced Oct 31, the Booking.com flights provider's MCP server (standard MCP over HTTP with OAuth 2.0, no SDK) covers real time search and fares across 550+ airlines in 75+ markets, plus seats, bags and passenger details in the conversation. Search and pricing are free once you request access; the traveller still pays on a hosted payment page unless you sign a commercial deal for in agent payments…

eGlobal Travel Media

Google opens a preview for plugging your remote MCP server into AI Mode and the Gemini app

The Connected Apps Platform preview, open for applications since Oct 30, lets you register a public remote MCP server (or Android AppFunctions) in Google Connected Console, test it in Gemini Developer Mode, and submit it for review to go live in Search AI Mode, the Gemini app, or both. Google tests every app before it ships and its policy bans tool descriptions written to steer the model toward your app over others (MCP guide)…

Google Blog

unslop is an open skill that strips the vibecoded look from AI-built UIs in Claude Code and Codex

New this week under CC BY 4.0, /unslop this dashboard audits a screen for purple gradients, glows, emoji labels, shouty copy and uneven spacing, proposes a direction from your product, then waits for approval before fixing it in nine passes from spacing to motion, on code or a live URL. It asks before any paid image generation, and can render a short before and after story video if you install Playwright and ffmpeg (walkthrough)…

GitHub (heydamir/unslop)

Tadpole is a from-scratch agent in 3,500 lines of Python, with LoRA adapters that teach Qwen3-4B tool use

Shown on HN Oct 31 under MIT, the framework free agent handles files, shell, web search and session compaction from a terminal or Telegram, on any OpenAI compatible API or a local Qwen3 4B, and its repo logs each build stage with tests and measurements. Its released adapters lift Qwen3 4B from 0.419 to 0.774 pass@1 on held out task types, but that's the author's own 140 task benchmark measured only inside Tadpole's harness (adapter)…

GitHub (shakewingo/Tadpole)

Archwhere puts every coding-agent PR on an architecture map that the agents keep up to date

Shown on HN Oct 31 under MIT, npx archwhere init drafts a one file architecture.json from your repo, a GitHub Action comments each PR with the components and security zones it touches, and archwhere check fails CI when files go unmapped, with check fix and a SKILL.md so agents maintain the map themselves. npx archwhere mission adds a local UI where component threads become GitHub issues your agent answers, and only you can merge…

GitHub (tapaskar/archwhere)

Spotted something we missed? Start a thread.