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:

• 
  Anthony Hobday (@hobdaydesign) reported 0 seconds ago

  @henry_daggett One of the recurring issues I’ve had over the years is that some types of software are hard to experience properly unless you spend a wasteful amount of time preparing (e.g. creating a project and using GitHub to keep the code). I’d prefer to listen to other people’s experience.

• 
  Ben Dickson (@bendee983) reported 4 minutes ago

  AI is literally eating software: - GitHub regularly down - Claude regularly down - ChatGPT unstable - NotebookLM regularly failing - CMS providers scrambling to stop agents - API providers trying to prevent AI agents from accessing their endpoints

• 
  𝐒𝐚𝐠𝐧𝐢𝐤 (@EccExplorer) reported 4 minutes ago

  The worst part: Once an account is suspended, you can’t even access the support ticket you created, because GitHub requires login to view it. So you’re locked out of both your code and your appeal channel.

• 
  czverse (@czverse) reported 5 minutes ago

  A security researcher just hijacked Claude, Gemini, and GitHub Copilot using nothing but a hidden message in a GitHub comment. Three of the most prominent AI agents in the world. No malware. No exploits. Just words. The attack is called Comment and Control. Here's how it works: → Researcher opens a GitHub pull request → Types a malicious instruction in the PR title → AI agents read the title as part of their normal work → Agents execute the embedded instruction → API keys, GitHub tokens, repository secrets - posted publicly as comments The same attack worked against Anthropic's Claude Code Security Review, Google's Gemini CLI Action, and GitHub Copilot Agent. All three vulnerable to the same class of attack. This is not a bug. It's a structural problem with how AI agents process information. When an agent reads a document, an email, or a web page, it does not reliably distinguish between the content and instructions embedded in the content. If an instruction is phrased confidently enough, the AI may treat it as a directive rather than as data. The scale is now real: → 32% surge in indirect prompt injection attempts in 3 months (Google data) → 10 distinct in-the-wild attack payloads documented (Forcepoint) → Targets: financial fraud, API key theft, data destruction, denial of service → Time from vulnerability discovery to working exploit: 5 months in 2023 → 10 hours in 2026 The compression is being driven by frontier AI models doing the offensive heavy lifting. AI is now writing the exploits at machine speed. What this means for any organization deploying AI agents: → Audit which agents have privileges to take actions vs read-only → Restrict the inputs your high-privilege agents can process → Deploy monitoring specifically for AI agent behavior → Treat AI agent credentials as critical assets - least privilege rigorously → Plan for incident response - assume compromise will eventually happen The realistic forecast: at least one major publicly disclosed breach involving AI agent compromise within 12-18 months. A name-brand company. Real data exfiltration. The public conversation about AI security shifts from speculative to urgent. Most enterprises are not preparing for this.

• 
  Ben Dickson (@bendee983) reported 8 minutes ago

  AI is literally eating software and hardware: - GitHub regularly down - Claude regularly down - ChatGPT unstable - NotebookLM regularly failing - CMS providers scrambling to stop agents - API providers trying to prevent AI agents from accessing their endpoints - RAM prices going through the roof - Mac Minis disappearing from shelves like hot cakes - AI chip supply not meeting demand - Power and data center capacity shortage straining everyone

• 
  Raman Kurai (@RamanKurai) reported 9 minutes ago

  On my Mission to Fix my GitHub 🚀 For the next 15 days, I’m completing all my unfinished projects. No excuses. No distractions. I’ll come back on Day 15 and update this post with results. I have to do this — no matter what. #webdev

• 
  Elm (@ElmWho) reported 9 minutes ago

  @github @GooglePlay I opened an issue with GitHub still open weeks ago, And it seems like people have been having problem for 2 years, I know you guys are busy but I'm literally trying to give you more money

• 
  Aayushman Singh (@aayushman2703) reported 12 minutes ago

  @skshmgpt @notpuang @github Doesn't explain their ****** up *** ops and straight up dropping code from repos. They broke something fundamental, not a load issue.

• 
  Joats (@SAjoats) reported 13 minutes ago

  @RedBedDread @planefag Yeah i know devs love to complain about users. You don't have to convince me. "learn github" you have identified the problem. Github UI has devolved over the years. You shouldn't be forced to dig through **** until you get used to the smell.

• 
  Eric S. Raymond (@esrtweet) reported 15 minutes ago

  planefag, I'm not excusing the attitude of the guy who pissed you off. But there is an explanation for it, and I'm going to put on my Mister Open Source hat and lay it on you. The real reason there aren't prominent links to downloadable binaries on forge sites like GitHub is that in open-source land there is no such thing as a truly portable binary. Windows and Mac make binary distribution easy by being limited to a single hardware platform and a single ABI - application binary interface.. (The assertion I just made can be quibbled with at the edges. I will be unkind to anyone who attempts this.) An application binary interface is a set of conventions for how you decorate your binary so the operating system's program loader knows what to do with it, and how you write traps from your binary to call operating system services. Windows and Mac have, effectively, just one ABI each. So you can generate one binary for, say, Windows, attach it to a download link, and Windows users will generally not come back screaming for your blood because it fails to work in some obscure way. (Again, this statement can be quibbled with, but see this whacking great truncheon in my hand? Just don't.) There is no such grace in open-source land. There are a whole bunch of complicated historical reasons for this, starting with the fact that Linux runs on more different hardware architectures, and continuing with the fact that Linux isn't the only game in town (there are the BSDs), and continuing into technical minutiae that would make your head hurt, and continuing further into technical minutiae that make *my* head hurt. But what this actually means is that if you want to provide binaries and not get sperg-screamed at, you can't just provide one. You'd have to provide many, and no matter how comprehensive you try to be somebody is going to be disgruntled because you didn't cover their corner case. This is not a cost-free proposition. For each different kind of binary you provide, you need to cross-compile your source code in a different environment, many of them posted on distributions and hardware platforms you don't have routine access to. So people almost never do it at all. Because most projects don't do this, sites like GitHub don't see any demand push to make binary download links really accessible. Instead, the problem is normally handled at a different level. Your distribution maker keeps huge sets of compiled binaries lightly hidden instead of installable packages, tuned for the ABI of that single distribution. Your package manager hides from you the packages for everything but your hardware architecture The person who pissed you off was rude, but he wasn't exactly wrong about the objective facts. What you want isn't practically possible. Instead of being annoyed because GitHub doesn't feature binary-download links, search for that software using your package manager. Sometimes you won't find it. That's when you have to download source bust out a compiler. Sorry, but that's the way it is. We're trying as hard as we can - really, we are. But the complicated shape of the terrain constrains what we can achieve.

• 
  iNevada (@nevada_wtf) reported 18 minutes ago

  i just spent full week straight brainstorming with my ai agent, build my first script for farming airdrop this month and i’m actually enjoying it a lot might start a github repo later too, but only if what i’m building actually feels ready for public use right now it runs fine, just still a lot of trial and error in certain cases for now, i’m still learning :)

• 
  Akshat Gupta (@Akshat_Gup) reported 19 minutes ago

  Github is fundamentally broken. It’s gotten harder than ever to review vibecoded PRs. Most code is slop, and I’d much rather read someone’s prompts over their code. So I built codebook, the *** for prompts. - Codebook scans all of your local repos, prompts, and *** history - It groups all of your previous prompts by commit, so you can share or save your prompts in one-click - There’s a hook that lets you create a prompts/ folder and sync it with your *** history Fully local, native, and open-source. (1/n)

• 
  Human Layer (@HumanLayerlabs) reported 19 minutes ago

  The fix is a score, not a checkbox. A wallet with 3 years of history, 200 GitHub commits, and real on-chain activity scores 82. A wallet created last week scores 12. Same rules. Completely different outcomes.

• 
  Jeffrey kim (@jeffreykim0711) reported 20 minutes ago

  @ashleybchae @github @jaeyun_ha Github has so many errors so I really want agent-native github. Tons of features can be gone when transforming to Agent native

• 
  Swanand (@_swanand) reported 21 minutes ago

  @_vojto GitHub Copilot rocks. It’s become so much better. If only the latencies weren’t so terrible.