Quick one today, but the lead story is a "stop what you're doing" kind of bug if you run OpenCode on Windows or WSL. Everything else is standards drama and a $65K excuse to vibe code a game. OpenCode's MCP sign-in bug let a server run commands on your machine Here's the setup: OpenCode connects to MCP servers to give your agent outside tools. When you sign in to one of those servers, the server gets to hand you a web address for OAuth — the "go here to authorize" link. On Windows and WSL, OpenCode passed that address into a PowerShell command without escaping it. Pillar Security found that a malicious address could carry a PowerShell payload and run it, with your access, the moment you signed in. The scary part isn't really the bug. Command injection happens. It's how it got fixed. Anomaly (the company behind OpenCode) patched it in version 1.18.33 on September 28, buried inside a pull request labeled as a "refactor" of browser-launch code. No security advisory, no CVE. Pillar reported it privately on August 29 and asked Anomaly to request a CVE; Anomaly's answer, per Pillar, was that a malicious MCP server is "outside our trust boundary." That means your dependency scanner and vulnerability alerts have nothing to match against. The only way to know you're exposed is to go check the version number yourself. Worth noting: your MCP server doesn't need to start out malicious. One you vetted and added months ago could get compromised later, or its domain could lapse and get bought by someone else, and you'd never know until sign-in runs a command on your box. The fix: upgrade to 1.18.33 or later if you're on Windows or WSL. If you can't upgrade today, run opencode mcp list, look at what remote servers you've got configured, and rip out anything you haven't fully vetted. macOS and Linux aren't affected — OpenCode doesn't route the browser launch through a shell there. And this is separate from the OpenCode Console breach reported around the same time, which may have exposed stored API keys.

Good news: shipping fast with AI is still the move. Bad news: "it runs" and "it's safe" are two different checklists, and most of us have only been doing the first one. The lead story: a production checklist for apps you vibe-coded SitePoint published a developer checklist this week for taking AI-generated apps from prototype to production, and the reason it exists is the stat buried in the middle: Veracode's 2025 GenAI Code Security Report found that 45% of AI-generated code samples failed security tests, tripping OWASP Top 10 issues. That's Veracode's own benchmark on code samples, not full apps or real breaches, so take it as a signal, not a verdict. Separately, a WIRED investigation from May found exposed data in apps built with popular AI coding tools in the wild. None of this means vibe coding is broken. It means the part where you turn a prompt into a login page and a dashboard was never the hard part. The checklist's actual advice is unglamorous: write down who can see what before you generate more features, ship one feature at a time instead of one giant prompt, and read the auth code specifically because a hidden button on the frontend does nothing if someone can just call your API directly. The example in the article is a classic: an app that lets a user view invoices needs to check that this user is allowed to see that invoice, server-side, every time. Not "logged in," but "logged in and allowed." If you've shipped something with Cursor or Claude Code and haven't gone back to check that every protected endpoint actually enforces permissions per record, that's probably this week's homework. [YOUR TAKE: have you run a security pass on something you vibe-coded, or found out the hard way?] Agent Skills just became official MCP plumbing Bloomberg helped standardize a spec called Skills over MCP, now an official MCP extension (SEP-2640), announced at AGNTCon + MCPCon Europe in Amsterdam. The short version: MCP servers can now serve both tools and the instructions for using them, as skills, from one place. An agent loads just the skill's name and description first, then pulls the full instructions only when it actually picks that skill. Update the skill on the server, every connected agent gets the new version automatically, no re-syncing files to every user. Bloomberg says it hit this problem building ASKB, its Terminal AI tool, once skills needed to scale past one team.

Good news if you've been staring down a TV that shows you an ad before it even lets you open Netflix: the fix might be a free AI agent and an afternoon. Bad news if you're into robots that text your friends by accident — more on that below. The lead story: people are vibe coding the ads out of their TVs A developer named Gathin T. got tired of his parents' Xiaomi TV shoving a sponsored ad in front of them before Netflix would even load. So he pointed Claude Code at the TV over Wi-Fi and let it work. The agent found the set, switched off 25 Xiaomi software packages, and built a new home screen that, in his words, "looks like an Apple TV now lol." He turned the process into an open-source project called cleanyourtv. As of Tuesday it had tested setups for 107 TV models across 37 brands, 10,700 GitHub stars, 1,200 forks, and 85 open pull requests. It works through Android's built-in debugging bridge (ADB), the same maintenance tool Android phones use, so nothing gets deleted — apps just get switched off, and a one-line script restores everything if you change your mind. Other people ran with it. Mihir Gupta swapped a 192 MB stock launcher for a custom 83 MB one. Ansub Khan said his parents' old Android TV got about 90% more free memory. Ryan Doyle sent Claude a product listing and asked "can you hack this?" and ended up with a screensaver tracking his Stripe payments. A separate project called dumbtv does the same job on Amazon Fire TV, killing about 65 packages — including the ad-tracking software — and replacing the launcher with a 41 KB app that has no internet permission at all. This isn't just a fun hack. Texas sued Sony, Samsung, LG, Hisense, and TCL in December 2025, alleging their TVs can screenshot your display every 500 milliseconds to feed ad targeting. Vizio already paid $2.2 million to the FTC back in 2017 for pulling viewing data from 11 million TVs without consent — and that didn't stop Walmart from buying Vizio for about $2.3 billion in 2024, partly for its ad business. The incentives for manufacturers to bloat your TV with ads aren't going away. The incentive to vibe-code your way out of it just got a lot more practical.

ForgeCode is telling everyone it beats Claude Code and Codex. The benchmark that's actually independently verified says something much smaller, and the real story buried underneath is about scaffolding, not which model is smarter. Also: Jev's two-week numbers landed, and WebWork gave assistants 120 ways to touch your timesheets. ForgeCode's benchmark win comes with a big asterisk ForgeCode, a terminal coding harness from Tailcall Inc., is topping TermBench 2.0 with an 81.8% score, and its marketing has run hard with that number as proof it beats Claude Code and Codex. One problem: TermBench 2.0 lives at tbench.ai, and ForgeCode submitted its own result. That's a self-reported score, not an independent one. Switch to SWE-bench Verified, the benchmark Princeton and UChicago actually run and hand-check, and the gap shrinks to almost nothing. ForgeCode paired with Claude 4 scored 72.7%, a couple points off top Codex and Claude Code setups. That's not the blowout the headlines suggested. The genuinely useful finding here isn't about ForgeCode winning or losing. Tailcall also showed that wrapping Gemini 3.1 Pro in ForgeCode's own orchestration layer, instead of a generic scaffold, pushed that same model's Terminal-Bench 2.0 score from 55% to 80.2%. Same weights. Same model. A 25-point swing just from changing how the tool plans and retries. Scale AI found the same pattern running three agent systems on identical Claude Opus 4.5 weights: scores ranged from 50.2% to 55.4% on SWE-bench Pro. The orchestration around a model moves the needle about as much as the model itself does. To be fair, ForgeCode isn't vaporware. Its GitHub repo has around 7,380 stars and 1,438 forks, with commits landing within the last day, and it's model-agnostic enough to run Claude, GPT, or Gemini depending on what you already pay for. That's a real project with a real community behind it. What's missing is any sign that working teams are actually ripping Claude Code or Codex out of production to replace them with it. Stars measure curiosity, not switched workflows.

Two stories today, same theme: AI is getting less chatty and more useful. Bluesky shipped a tool that turns prompts into shareable mini-apps, and the "model that can't write a sentence" just posted the fastest adoption numbers Vercel has ever logged. If you're building with Claude Code or Cursor, both are worth five minutes. The lead: Bluesky's Attie turns prompts into shareable mini-apps Bluesky announced at TechCrunch Disrupt that Attie, its AI feed-builder from earlier this year, now lets anyone build websites, games, polls, and dashboards from a prompt and share them with other people. CEO Toni Schneider's pitch: "social vibe coding or vibe coding for your community." Translation: instead of solo vibe coding a thing only you'll ever open, you build something your group actually uses. The mechanics are simple. You type a prompt, Attie generates the app, and it gets a shareable link on the attie.site domain with your attribution at the bottom. The Verge's reporter made a Pokémon poll in a few minutes, and the showcase page has a Flappy Bird clone, a drum machine, a quiz, and an idle clicker. These are lightweight novelties, not production apps, but that's the point: distribution is baked in. Apps run on ATProto, the same protocol under Bluesky, and COO Rose Wang's framing is the real hook: "Who's using the thing you're building, because we haven't solved that problem with distribution." Attie comes with Bluesky's audience attached instead of you having to go find one. Worth noting: Bluesky's own active user base has reportedly been shrinking even as the company expands into tools like this, so "distribution" here means a network that's not currently growing, not a guaranteed audience. [YOUR TAKE: worth trying Attie this week, or is "vibe coding for your Discord" already solved by other tools you use?] Quick hits GitLab's MCP server goes GA, and it covers the whole lifecycle GitLab 19.5 ships a generally-available MCP server that lets Claude Code, Cursor, Copilot, Codex, Gemini, and other MCP clients work across code, merge requests, work items, and pipelines, under your existing GitLab permissions. New tools include semantic_search (find code by meaning, not filename), get_pipeline (see why a build is red without digging through logs), and the ability to kick off full Duo Agent Platform flows without leaving your editor. Rate limits on GitLab.com: 60 calls/min on Free, 600 on Premium/Ultimate. If your agent's been guessing at GitLab's REST API and burning context doing it, this is the fix.