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:

• 
  Debbie O'Brien (@debs_obrien) reported 3 minutes ago

  I promise I am not being paid to say this but @cursor_ai mobile experience is just amazing. Why are more people not talking about this and sharing it? Why isn’t their a dev rel team on this? Today while nap trapped (in car while boys were napping) I decided what if I tried the Cursor app and fixed the login for my Playwright demo site. one prompt (i know thats what they all say but it was ) Cloud agent spun up. Live updates on my phone meaning when it was ready to review I could see or if it was still working. That meant I could continue doing other stuff and not have to keep the app open. Merge straight from the app or jump to @github app to review and merge. It was so effortless that I started giving it more stuff to do. At this stage while I was preparing food. Improve mobile view on site. Just reviewing that now but look at the screenshots built right into the app to show me what the agent had done. Amazing. Im not saying we should all work weekends and always while we do other stuff but when you have open source repos to maintain and don’t have time to dedicate to them well this gives you a way to do it. Its a fantastic experience and makes me want to fix more stuff. Now thats just me at the weekend with a few mins to spare. Imagine what you can do at corporate level, sending tasks off to cloud agents and getting feedback on your phone while traveling, having a coffee or whatever, easily seeing when something needs your attention and acting right away making progress move much faster. This is amazing. By far the best cloud agents experience. Let me know if I am wrong and there is a better one out there as would love to try that out but I believe Cursor is winning this one hands down. If only more people knew about it….

• 
  Spyros (@SpyrosOnX) reported 3 minutes ago

  @GitHub account suspended – one week has already passed. No notification, nothing strange in my repos (95% private). Paying customer. No response after raising a ticket and contacting @github on X. Locked out of my code/issues and third-party sites using GitHub login. Thanks @Microsoft

• 
  David putra (@davidputra2112) reported 3 minutes ago

  seven GitHub repos pushed in four days, a working USDC escrow contract, and a README that flat out says "don't put real money through this yet, we're not audited." that combination is rare enough that I sat down and actually read the whole thing. $h3gt 4eYp69P1VU946efStVzV41gQvYjMucpcuYEbofc6pump 👇

• 
  PatriotOfTexas (@FreeTXPatriot) reported 4 minutes ago

  @KanekoaTheGreat When I first started using Ai, GitHub copilot. I did 6 months of work in 5 days.. now, I still routinely solve issues in an hour or 2 that would have taken a week or 2

• 
  Will Stith (@TheStithLord) reported 4 minutes ago

  @sbilstein Please! Save us from GitHub. I work in a tiny startup so I personally don’t run into performance issues, I just can’t stand the UI/UX of GitHub and want to move off asap.

• 
  ZomboDB (@zombodb) reported 4 minutes ago

  Whenever GitHub is down I smoke meat.

• 
  Abbas Asadzade (@AbbasAsadzade) reported 10 minutes ago

  The new wave of AI-native business frameworks all end with the same step: compound into a skill and rerun on cron. Looks clean on GitHub. The part that actually decides whether it works is what happens when the skill is slightly wrong and keeps compounding the error every night while nobody’s watching.

• 
  Miguel Sanchez (@MSanchezWorld) reported 12 minutes ago

  @sama I'm sure you guys have thought of this but just in case why not make it so you can only do security checks if the company validates its ownership of the domain, server, and GitHub? This way we can all start hardening our software for the impending hack apocalypse. I have many more good ideas like this if you want to hire me. LOL

• 
  J Filipe (@jrmromao) reported 12 minutes ago

  GitHub Actions hit another outage yesterday, partly due to "surging AI usage." And 25% of businesses are already delaying AI projects over costs. This isn't just about efficiency anymore; it's about stability and project survival. We have to get AI spend under control.

• 
  Steven Anglin (@StevenAnglinn) reported 15 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.

• 
  Eric (@transurfer) reported 16 minutes ago

  Ever seen a tool that actually gets better fast? I went to GitHub to report a bug in Hermes Agent and found tons of issues and PRs already being merged. Looking forward to what's next.

• 
  VardhanInsights (@harsh_sing91766) reported 17 minutes ago

  India’s I4C orders GitHub to take down Jack Dorsey’s Bluetooth-based app Bitchat, citing security risks. The tool, used by Delhi protesters during internet shutdowns, now faces the same fate in India as in China. #CyberSecurity #TechPolicy #StudentProtests

• 
  Mo Syed (@msyed_) reported 17 minutes ago

  Tokenmaxxing is dead. Long live useful AI. There’s a new productivity cult in town. Use more tokens. Run more agents. Burn through as much compute as possible. Somehow, the company that spends the most on AI wins. That idea is starting to look a lot less clever. More tokens can mean better results. More agents can mean more work gets done. But only up to a point. After that, you’re just paying AI to walk in circles. The bottleneck usually isn’t the model If your organisation has unclear goals, bad processes, weak data, or nobody checking the output, throwing more tokens at the problem won’t save it. It just produces expensive confusion faster. There’s also a small conflict of interest here: model providers sell tokens. So naturally, some of the advice coming from the industry sounds a bit like: “Use more of the thing we charge you for.” It’s the AI version of a mechanic recommending an oil change every 3,000 miles, or a toothpaste ad showing someone covering the entire brush. The product is useful. The suggested quantity may be doing a little too much work. The smarter rule: measure the bill before you scale Once an AI application becomes more than a basic experiment, instrument it. Know what it costs to run. Maybe one query costs 50 cents. Maybe a ten-minute conversation costs $3. Those numbers make the conversation much more useful: Is the task worth automating? What happens if usage grows tenfold? Which model is good enough? Where does human review still make sense? You don’t need a 40-page business case. You just need to know whether your “efficient” AI workflow is quietly setting money on fire. Don’t marry your model provider The other important rule is simple: Keep your options open. Even in an early prototype, design things so you can switch between providers. That might mean supporting several APIs, testing open-weight models, or keeping the model layer separate from the rest of the application. Because today’s best model may be tomorrow’s overpriced legacy system. The winning companies won’t necessarily be the ones using the biggest model. They’ll be the ones that can change models without rebuilding everything from scratch. DeepSeek just made “small model” look embarrassing DeepSeek’s updated V4-Flash-0731 reportedly overtook its own larger V4-Pro model on independent tests. Same basic architecture. Better fine-tuning. The smaller model scored 50 on Artificial Analysis’ Intelligence Index, just behind GPT-5.6 Luna at maximum reasoning. It also beat its earlier preview on agentic coding tasks, reaching 82.7% on Terminal-Bench 2.1. And the price is the part that makes this genuinely interesting: $0.14 per million input tokens $0.0028 per million cached input tokens $0.28 per million output tokens The model is also available under an MIT licence, and a quantised version can run on a machine with around 110GB of memory. That puts capable AI within reach of teams that don’t want to send everything to a cloud API. The AI model race is becoming a cost race The biggest model used to win the conversation. Now developers are asking a more practical question: “Can this model do the job cheaply enough to run all day?” That matters because agents are greedy. They read files, call tools, retry failed actions, check their work, and start again. A model that costs 50% less per task can turn an impressive demo into a viable product. Bug triage. Invoice reconciliation. Customer support. Internal research. The economics are changing underneath all of them. Claude just helped break a post-quantum encryption candidate Anthropic’s Claude Mythos Preview found a weakness in HAWK, a proposed digital signature scheme being considered by NIST for post-quantum cryptography. The result? HAWK was withdrawn from the competition. The important detail is that this wasn’t a movie-style “AI cracks the internet” moment. The attack still required expert direction, multiple agents, working code, and around 60 hours of effort. It also cost roughly $100,000 in API usage. But the model found a weakness that had survived years of expert review. That’s the part worth paying attention to. AI doesn’t need to invent brand-new mathematics to be useful in cybersecurity. Sometimes it just needs to combine known techniques more patiently and thoroughly than a human team had time to do. Better lockpicks can help build better locks The same technology that finds flaws can help prevent them. Researchers created SecureForge, a system that automatically improves an AI coding assistant’s system prompt to reduce security vulnerabilities. Simply telling a model to “write secure code” didn’t work very well. SecureForge tested generated code, identified vulnerabilities with static analysis, and then improved the prompt based on what went wrong. Across the models tested: SecureForge produced vulnerable code 11.8% of the time A normal “write secure code” prompt failed 20.1% of the time That’s not perfection. But it’s a useful reminder that secure AI coding needs more than good intentions and a sentence saying “please avoid SQL injection”. The new code models are about to learn from AI-written code Hugging Face released The Stack v3, a huge new dataset built from public GitHub code. The training set contains: Around 4.9 trillion tokens 15.9 terabytes of filtered code 713 programming languages Code from roughly 173 million repositories The major upgrade is that it preserves whole repositories, not just isolated files. That matters because modern coding agents need to understand how a project fits together. They need to trace dependencies, follow functions across files, and work with the structure of an entire codebase. But there’s an odd loop forming. The dataset includes code written or assisted by today’s AI tools. So the next generation of coding models may learn partly from the output of the previous generation. AI is starting to train on its own footprints. The big takeaway The AI industry is moving from: “Use the biggest model and burn as many tokens as possible” to: “Use the cheapest system that reliably solves the problem.” That means measuring costs, keeping model providers interchangeable, improving security prompts, and giving agents enough context to do real work without letting them run wild. Tokenmaxxing was a fun slogan. Operational discipline will pay the bills.

• 
  Carat (@Caarat1) reported 18 minutes ago

  He's 24. He built the LiDAR replacement Tesla just paid $3.8 million for on six broken iPhones he pulled off eBay for $40 apiece and a Raspberry Pi 5 he wired together in a Boulder apartment Each of the six iPhone X TrueDepth sensor modules is the same infrared dot projector Apple ships in every FaceID iPhone - the exact same hardware that maps a human face at 30 frames per second in complete darkness. He mounted them in a hexagonal array on a $8 3D-printed bracket he printed at a Boulder makerspace, wired them to the Pi through custom USB adapters he soldered himself, and wrote a Python driver that stitches all six depth streams into a single point cloud at 34 frames per second. Effective range: 40 meters. Total hardware cost: $240. A single Velodyne HDL-64E LiDAR unit that Tesla previously benchmarked against costs $75,000 He posted a demo video to GitHub in October showing his rig tracking pedestrians, cyclists, and parked vehicles across a Boulder intersection with 94% accuracy against Velodyne ground-truth data. A Tesla Autopilot engineer found it through a Hacker News thread three days later. Elon Musk quote-tweeted the demo the following Saturday - called it "the sensor architecture we should have shipped in Hardware 4 instead of paying Mobileye a decade of licensing fees." By December Tesla had wired $3.8 million into his account for the sensor fusion algorithm and a three-year consulting contract to integrate the TrueDepth array pattern into Autopilot Hardware 5 rolling out to every new Model Y this quarter Tesla runs at a $900 billion market cap running Autopilot on the premise that autonomous vehicle perception requires their proprietary silicon and their multi-billion-dollar sensor supply chain. Mobileye sits at $15 billion in market cap selling ADAS chips on the same premise. Elon Musk just paid a 24-year-old in a Boulder apartment more for a Python driver and six broken iPhones than most Tesla Autopilot engineers earn in a decade

• 
  Big bro (@reactreaper) reported 24 minutes ago

  @dantechceo i mean logging in with gmail triggers a otp sent to gmail for login i think that's enough for me to not touch it given that i get logged out every few days, while i don't even remember the last time i logged in to github