cuppa

livebrewed 03:04 AM
← Jun 22

Tuesday

jun23

2026

Jun 24 →

the brief

Today’s pulse leans practical: Claude’s ecosystem tightens with CLI-based MCP auth and a one-click Claude Design→Vercel deploy. Open models advance with PP‑OCRv6, while papers probe FP4 shrinkage bias and “role confusion” prompt injection. Cloudflare’s deep dive into a long-standing hyper bug and a timely AI security conversation round out a developer-centric slate.

the poursit · sip · 8 items

pulse

(03)
  • anthropics/claude-code· First-partyJun 22, 08:37 PM

    Claude Code adds CLI MCP auth

    Adds mcp login/logout CLI for non-interactive server auth, workflow filters, a Skills tab, iTerm2 teammate mode, and AWS credential refresh quality-of-life.

    v2.1.186 — What's changed Added claude mcp login <name> and claude mcp logout <name> to authenticate MCP servers from the CLI without opening the interactive /mcp menu, with --no-browser stdin redirect support for completing over SSH Added status filtering (press f) to the /workflows agent detail view Added a "Skills" section to the /plugin Installed tab Added teammateMode: "iterm2" setting with a warning when auto mode cannot find the it2 CLI Added "Claude Platform on AWS - refresh credentia...

    signal 9hype 1changelogclaude_codemcptechnicalsource ↗
  • vercel/news· First-partyJun 23, 12:00 AM

    Deploy Claude Design directly to Vercel

    Claude Design can now deploy designs as live Vercel projects via a built-in destination and MCP server integration, returning shareable URLs without leaving the canvas.

    Deploy from Claude Design to Vercel — Vercel is now a send-to destination in . When you finish a design, you can send it to Vercel and get a live URL back without leaving your canvas.Claude Design Claude Design deploys the design as a new project in your connected Vercel account and returns a URL you can open and share. When you're ready to share a design, add Vercel as your destination in the 'Share' menu and connect the to get started.Vercel MCP server about using Claude Design and Vercel t...

    signal 7hype 2vercelclaude_designmcplaunchsource ↗
  • huggingface/blog· First-partyJun 22, 01:18 PM

    PP-OCRv6 lands on Hugging Face

    Lightweight multilingual OCR models (1.5M–34.5M params) support 50 languages, enabling on-device and server OCR pipelines with open weights and ready-to-use demos on Hub.

    PP-OCRv6 on Hugging Face: 50-Language OCR from 1.5M to 34.5M Parameters

    signal 6hype 1model_releaseocrmultilinguallaunchsource ↗

findings

(03)
  • cloudflare/blog· First-partyJun 22, 06:00 PM

    Cloudflare uncovers long-standing hyper bug

    Reworking Workers Images surfaced a cross-version bug in Rust’s hyper HTTP library; the write-up details reproduction and upstream impact, a must-read for Rust web services.

    How we found a bug in the hyper HTTP library — By rearchitecting the Images binding, we accidentally uncovered a bug that existed in the open-source hyper library across multiple major versions.

    signal 7hype 1bug_reportpostmortemrusttechnicalsource ↗
  • simonw/blog· AnalysisJun 22, 11:59 PM

    Prompt injection as role confusion

    New paper reframes prompt injection as role misassignment in agent systems, with a readable companion blog that clarifies attack surfaces and suggests actionable mitigations.

    Prompt Injection as Role Confusion — <p><strong><a href="https://role-confusion.github.io">Prompt Injection as Role Confusion</a></strong></p> First, I absolutely love this:</p> <blockquote> <p>This is a blog-style writeup of the paper.</p> </blockquote> <p>I wish <em>every paper</em> would come with one of these. Academic writing is pretty dry - the impact of a paper can be so much higher if you publish a readable version to accompany the formal one.</p> <p>Charles Ye, Jasmine Cui, and Dylan...

    signal 7hype 1paperprompt_injectionagent_securitytechnicalsource ↗
  • pplx/tech-writeups-24h· ResearchJun 23, 03:00 AM

    UFP4 recipe fixes FP4 shrinkage bias

    ArXiv study traces geometric origins of FP4 shrinkage bias in LLM pretraining and proposes an Unbiased FP4 recipe that stabilizes training with competitive downstream performance.

    Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe — Abstract page for arXiv paper 2606.20381: Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe

    signal 6hype 1paperquantizationfp4technicalsource ↗

voices

(02)
  • simonw/blog· AnalysisJun 22, 11:43 PM

    Porting Moebius inpainting to the browser

    Simon Willison documents bringing the 0.2B Moebius inpainting model to run client-side with Claude Code assistance, showing practical steps for WebGPU-ready, local image edits.

    Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code — <p>This morning <a href="https://news.ycombinator.com/item?id=48630171">on Hacker News</a> I saw <a href="https://hustvl.github.io/Moebius/">Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance</a>, describing a small but effective inpainting model - a model where you can mark regions of an image to remove and the model imagines what should fill the space. The released model <a ...

    signal 9hype 1inpaintingbrowser_inferenceclaude_codetechnicalsource ↗
  • latentspace/podcast· AnalysisJun 22, 09:06 PM

    Red-teaming after Mythos, explained

    Zico Kolter and Matt Fredrikson argue AI security is a distinct discipline from cybersecurity, outlining threat models and org practices for post-Mythos model deployments.

    Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan — OpenAI boardmember Zico Kolter and Gray Swan CEO Matt Fredrikson join swyx to explain why AI security is not just “cybersecurity with AI”

    signal 5hype 2ai_securityred_teamingpodcastculturalsource ↗