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:

• 
  Draphilo (@Draphilo) reported 42 seconds ago

  @techtasium GitHub was down before?

• 
  Xah Lee (@xah_lee) reported 5 minutes ago

  August 17, 2026 github down for 8 hours. whats the cause @grok

• 
  Simplifying AI (@simplifyinAI) reported 6 minutes ago

  Boris Cherny is not gonna like this.. Researchers at the University of Luxembourg analyzed real production telemetry from coding agents, Claude Code, Cursor, GitHub Copilot, to see how often these agents ask for a second page of results when a tool response gets cut off. The answer, across the entire dataset: zero times. When a search returns too many results, the agent works with whatever made it into that first batch and moves on, even when what it needed was one page later. • 37% of "get epics" calls exceed the token budget in production • 28% of "get merge request diffs" calls exceed it too • 0% of truncated responses were ever followed by a second-page request • A basic keyword ranking raised the odds of the right item landing in that one page from 24.2% to 35.0% • Fancier ranking signals on top made it worse, not better Noisy tool output isn't just a cost problem. It's a silent correctness problem too, whatever doesn't make the first page effectively stops existing to the agent. wrote the full guide on keeping that output from piling up in the first place. full article below 👇

• 
  Jarkko (@JarSyr) reported 9 minutes ago

  Then SpaceX acquired Cursor on August 14. Three days later, Cursor launched Origin, its own GitHub competitor. And two weeks after that, OpenAI publicly said it wasn’t confident SpaceX would follow its terms and started winding down the relationship. Hard to ignore the timing.

• 
  Ashutosh Kumar (@akx_build) reported 10 minutes ago

  Everyone dunks on @github when it goes down. Fair. Outages suck. Also true: they still give away an absurd amount of infrastructure, at a scale almost nobody else even tries to match

• 
  Kyronis (@kyronis_talks) reported 14 minutes ago

  Your model isn't guessing when it lies to you. Someone measured it. 720 fabricated numbers pulled out of a language model. 565 of them came out at near zero entropy. That's 78%. The model was maximally confident while inventing a number. No hedge, no wobble, nothing in the probabilities to catch. It genuinely did not know it was wrong. Which means every "just check the model's confidence" hallucination filter is blind to four out of five fabrications. That number comes from SIMURG, and the guy who found it published it against his own product. His first version read logprobs. His own data said logprobs are not enough. So instead of burying that he shipped three more layers on top and put all of it on GitHub for free. It now resamples the claim, checks it against Wikipedia and live web results instead of just asking the model again, and when the sources disagree with the answer it abstains instead of asserting. And it learns while it runs. Every thumbs up or down from a real user is one training step. No retraining window. An AI that says "I don't know" because a search result told it to.

• 
  StonedModder (@StonedModder) reported 14 minutes ago

  @SeregonWar @github GitHub uses AI for minor things. They actually have a trust and safety team. Issue is they’ve laid off so much of that team tickets are slooow

• 
  orangesurf (@OrangeSurfBTC) reported 15 minutes ago

  @callebtc I wouldn't update to unreproducible binaries as an emergency fix - a project maintainer's github/signing accounts could be compromised and you end up installing malware. Given this, I wanted to verify who had signed off on the v26.06.7 release

• 
  Ryan Shaw (@ryankshaw) reported 16 minutes ago

  That thing where GitHub is always down now because of increased load from massive agentic activity, is just a preview of what is going to happen with almost every service we depend on. If it does something useful, and has an API, or a scrape-able website, it's gonna get hammered

• 
  Mallik Cheripally (@heyiammallik) reported 17 minutes ago

  A maintainer published a security fix in a public PR. About 10 minutes later, his live server started receiving requests trying to exploit the same weakness. He wanted to know how easy it had become to turn a small hint into an exploit. So he gave an AI only a rough description of the bug. In under a minute, it found related issues and built a working exploit locally. We don't know whether AI sent those requests. His test still showed that a rough description may be enough. Once a security fix is public, the patch can tell an attacker where the bug was and how to recreate it. I think the release should be ready before the PR goes public. Test the fix privately. Prepare the update. Let an agent search for similar bugs. Then publish the PR and release the update together. Many open-source maintainers cannot do this cleanly today. Private forks can lose the CI, bots and reviewer access they normally use. GitHub needs a better security fix mode. Keep the work private while it is tested, with the same CI and reviewers as a normal PR. Make it public only when the fixed version is ready to install

• 
  Will (@inovercy83) reported 18 minutes ago

  @GithubProjects If you still use github 2026 you are wrong. Just say "Calude Publish, if someone finds a bug fix it"

• 
  Sentient (@sentient_agency) reported 20 minutes ago

  I cancelled YouTube Premium last week. The thing that replaced it costs nothing, runs in any browser, and was built by 361 volunteers on GitHub. Invidious does the four things people actually pay Premium for. No ads. Background audio on mobile. Watch without an account. Privacy from Google's tracking. All of it, free. The part that surprised me most is how clean the experience is. The entire page renders without JavaScript. Pages load in milliseconds because there's nothing to load. No tracking pixels. No autoplay traps. No recommended-for-you algorithm trying to swallow your evening. You can: > Subscribe to channels without a Google account > Get notifications when they post > Import your full YouTube subscription list in one click > Switch between dozens of public instances if one goes down > Self-host it on a $5 VPS if you want full control > Use it with the Privacy Redirect extension to auto-redirect every YouTube link The repo has been actively maintained for years. Latest release was February 2026. (100% Opensource and free to use)

• 
  Enjoyer (@Enjoyer100x) reported 21 minutes ago

  @StandartXBT Brother, That github Ozzy posted, was made 4 days before Pons V2, they switched to it because they wanted to keep what was built for V2 private. It’s that simple. Either way, PePons is a strong meme and the team running it are doing a great job, perhaps they thought the info was true rather than larp. They are also in talks with Ozzy and he has been helping them transistion to V2. Why would he do that for larpers. People like you think you are doing good for the space, but actually you are part of the problem. If you can do true research the don’t speak.

• 
  Massimiliano Sermi (@MaxSermi) reported 25 minutes ago

  Imagine I tell an AI coding agent: “Fix the pagination bug in the customer dashboard.” The agent has GitHub credentials that technically let it modify much of the repository. But what authority did I actually delegate? #CodingAgents #SoftwareEngineering #DeveloperExperience

• 
  Louround (@Louround_) reported 28 minutes ago

  Nvidia is paying $13b for HuggingFace (3x last valuation) because the real issue in robotics isn't compute or hardware but data. Every robot needs millions of human demonstrations to learn how to move, and that data is spread across labs with no standard. HuggingFace became a leader with Lerobot (the github of robotics data) and Pollen robotics, Nvidia already owned the sim, models and chips but buying HF gives them the full stack with data → training → deployment. solana:69LjZUUzxj3Cb3Fxeo1X4QpYEQTboApkhXTysPpbpump is building the same coordination layer for VLA agents, but on chain and permissionless, Simarena for testing, Foundry for training and Machinefun for incentives. Same thesis Nvidia just validated with $13b but solana:69LjZUUzxj3Cb3Fxeo1X4QpYEQTboApkhXTysPpbpump is at $4m fdv. ⏳