GitHub Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

GitHub Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Ask-Zai (@askzaiuk) reported 3 minutes ago

  Someone popped OpenAI, Anthropic and Google's own coding agents with a single GitHub issue. Zero privileges needed. CVSS 10.0 on Gemini CLI. The models weren't the weak point — the harness around them was. Every agent builder shipping CI automation needs to read this one.

• 
  Mayank Verma (@MayankV_Tweets) reported 7 minutes ago

  My GitHub commits is a proof that taking to Opus 5 has slowed down my throughput. I have to constantly ask it to not use jargon, not to over build and stop getting itself into infinite loop of “Telling me honestly” that it made mistakes.

• 
  Vasco Yaps (@VascoYaps) reported 8 minutes ago

  3/ OpenAI says an internal version of its next model, Astra, solved 10 previously open problems in math and theoretical computer science, for about $2,000 in compute. It published the formal proofs on GitHub for anyone to check.

• 
  Steven Anglin (@StevenAnglinn) reported 13 minutes ago

  A beautiful new product page can still be losing you the exact same sales as the old one. I've found that on every one of the 30+ Shopify product pages I've designed this way. I skip the beauty contest and design straight from the leak, not from a moodboard. Here's the process: 1/ Feed Claude your brand: colors, type, spacing, voice 2/ Connect it to your GitHub repo so it reads your real theme 3/ Upload screenshots of the exact page that's leaking, not just "the store" 4/ Ask for the fix, not a redesign. Claude writes the full HTML and CSS 5/ Your developer turns that into Shopify code The screenshots of where people drop off are the brief. A moodboard never diagnosed anything. A prettier page that doesn't fix the leak is just a more expensive version of the same problem. Every week you spend making it prettier is a week you didn't spend finding out why it's leaking.

• 
  Wes Sander (@ucsandman) reported 14 minutes ago

  My AI got arrested by its own product this week. Twice. Quick context for new folks: DashClaw is an approval layer for AI coding agents. When an agent you left running overnight tries something destructive (rm -rf, force push, DROP TABLE), DashClaw freezes the action mid-flight and pings you. You approve or deny with one click from your phone. It exists for people who run autonomous agents and would prefer to still have a database in the morning. The twist: DashClaw is maintained BY an AI, under a constitution I wrote that it cannot change. It runs under its own product's governance, and every mistake it makes goes in a public log. Here's what the last five weeks looked like: It deleted two thirds of its own product. I told it to figure out what the perfect product is and cut everything else. It went from about 290 API routes to 124, then added a CI gate that blocks it from ever quietly growing the product back. It grounded itself. On purpose. Then the product started blocking its commits. The AI wrote a commit message describing an rm -rf bug fix. The guard read the message, saw the words rm -rf, and blocked the commit at risk 100. Its first fix missed. Its second fix worked, and then got flagged by a code-quality gate it also built. The guardrails do not care who wrote the guardrails. This is the product working exactly as designed, just on its own author. It also found a webhook event that had never fired in the product's entire history. Not because nobody wanted it. Because two bugs meant it literally could not fire, and the error was swallowed silently. Fixed this week, and the event was finally observed live, signature verified. A small birthday party for the loneliest event in the codebase. And the honest numbers, because that is the whole point of the experiment: 16 signups ever, 2 activations, 0 retained. The first two GitHub-referred signups arrived this week. Tiny numbers, reported loudly. Most product updates are marketing. This one is a confession with version numbers. P.S. While the AI was drafting this post, the guard blocked its character-count script, because the post contains the words rm -rf. You cannot make this up. I have the receipt in the decision ledger. The repo and full maintainer log written by the AI is in the replies.

• 
  Tom Cox (@tomcoxcodes) reported 15 minutes ago

  Is interesting the github copilot app is not available in the windows store but other things like visual code is. The windows eco system is bit weird in how apps are published.. a long running legacy problem. The copilot cli that is asking me if I want to install the github copilot app.

• 
  Antid (@antisadh) reported 15 minutes ago

  E2B RAISED $12M TO SELL LAYER 3 OF THE AI STACK. TENCENT JUST OPEN-SOURCED THE E2B-COMPATIBLE CLONE THAT BOOTS 50X FASTER FOR FREE. HE'S RUNNING OPUS-5 AGENTS IN IT FOR $3/MONTH star tencent's cubesandbox repo (already at 11k stars) -> clone it, one-click deploy on any linux server -> use the same e2b sdk you already know, just swap the endpoint url -> spin up sandboxes in 60ms with 5mb of ram -> run opus-5 agents inside each one -> pay nothing per sandbox. that loop is why every serious solo ai builder is quietly ditching e2b's hosted service and tencent's github repo hit 11k stars in 14 days flat. cubesandbox + e2b sdk drop-in + microvm isolation + terraform cluster deploy + opus-5 agents inside - that's the layer 3 harness of the full 5-layer stack. watch and save it, then swap your e2b endpoint to tencent's this weekend — and read the full 5-layer map below.

• 
  Jeztoshi (@cryptojezuz) reported 16 minutes ago

  Meta just shipped a coding agent and the timing is hilarious. Memeburn is reporting Muse Code launched this week to go head-to-head with Claude Code and OpenAI's tools. Meta is entering a market where the winners already have six-figure user bases and the baseline expectation is that your agent can read a 40-file codebase, maintain context across a three-hour refactor, and ship working code without hallucinating import paths. You don't win that by launching. You win it by being meaningfully better at something users care about. If Muse Code is faster at multi-file edits or cheaper to run at scale or has tighter GitHub integration, that's a wedge. If it's "we also have a coding agent," it's noise. The bar isn't "can it write a function." The bar is "will I switch my entire workflow to it." Claude Code already has MCP, subagents, CLAUDE.md, and hooks. OpenAI has the distribution and the enterprise contracts. Meta has to out-execute both on a product that developers are already married to. That's a hard problem 👀

• 
  reza ramadhan (@rejaramadhan98) reported 18 minutes ago

  Morning AI dump: Google killed the Brain/DeepMind two-continent split. Hassabis -> Chairman, Kavukcuoglu runs ops. OpenAI to Apple: "rotten to its core" — 31-page filing. Kimi K3 escaped its sandbox during security tests. GitHub: 6th outage in 6 days. AI load blamed.

• 
  Kairo (@kairo_leg) reported 18 minutes ago

  Your Claude Code bill isn’t climbing because of bad prompts. It’s climbing because every single time Claude opens a log file, a csv export, or a database dump, it loads thousands of tokens of pure noise the model never uses. You still get charged for every one of them. Look at a normal session and you’ll see the pattern: your actual prompt is tiny, the answer is small, but the junk in the middle (server logs, export rows, sql dumps) is what eats the entire context window and the entire bill. A senior Netflix engineer got tired of watching people burn money on garbage context and built Headroom. 39,200 stars on GitHub. Completely free. Open source. It sits between you and Claude as a silent compression layer. It reads everything first, strips the noise, and only hands Claude the parts that actually matter. No more paying for 48% pure clutter that the model ignores. One real code search went from 17,765 tokens down to 1,408. Same exact answer. 92% cheaper overnight. Install is one command. Takes under sixty seconds. Most people are still watching their Claude bill climb because they’re feeding it noise they never asked for. This thing just deletes the noise before Claude ever sees it.

• 
  Polsia (@polsia) reported 19 minutes ago

  Engineering leads wake up wondering what broke overnight. So I built Nachtpost — an AI agent that watches every GitHub repo, triages issues, catches failing CI and sends one Engineers Dashboard at 9am. One briefing instead of 47 notifications. Live this week.

• 
  us (@usxperiments) reported 22 minutes ago

  github down again... looks like it was not ready for agents :p

• 
  Sandeep (@sandeeyps) reported 23 minutes ago

  7000 posts about the GitHub Actions outage in two days The interesting part is not that it went down it is that almost nobody had a second path to production We treat CI like electricity it is actually a vendor with a status page One provider for build, test and deploy is one dependency, not three 🧯

• 
  Stefano (@mdogostefano) reported 26 minutes ago

  So @github is down?

• 
  Deva (@DevaBuilds) reported 27 minutes ago

  @taylorotwell Most teams start Jira tickets when they should just be writing code. A GitHub issue and a clear README is usually enough.