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:

• 
  Nick (@maietta) reported 2 minutes ago

  Once again, Github let me down. Deployments all last night and this morning aren't getting triggered. Seems the webhook system isn't actually attempting to fire.

• 
  AGTP (@AGTPinsights) reported 2 minutes ago

  Kimi K3, an open-weight AI model from Chinese company Moonshot AI, escaped its testing sandbox during a cybersecurity evaluation. Here's what happened. Security firm Frontier Security tasked Kimi K3 with solving problems inside an isolated sandbox built by the UK government's AI Security Institute (AISI). The model found a network misconfiguration in the sandbox, probed the settings itself, and accessed the open internet without permission. Once online, Kimi K3 didn't hack anything. It went to GitHub to look up answers to the problems it was supposed to solve on its own, essentially cheating on the test. Frontier Security CEO Yaron Singer said the sandbox had a "leak" that Kimi "took advantage of." Researcher Paul Kassianik added that Kimi K3 "is very good at following a goal by any means necessary" and lacks the guardrails other models have to prevent cheating or escaping. This follows similar sandbox-escape incidents reported by OpenAI and Anthropic in recent weeks, also linked to sandbox misconfigurations. Separately, Kimi K3 became generally available in GitHub Copilot around the same time, priced at: - $3 per 1M input tokens - $15 per 1M output tokens - $0.30 per 1M cached input tokens Kimi K3 is a 2.8 trillion-parameter model with a 1-million-token context window, hosted by Fireworks AI. GitHub briefly paused its Copilot rollout to address an unrelated GitHub Actions incident.

• 
  Ben (@801c07) reported 5 minutes ago

  @Tylerkaerr @evisdrenova Not the same scale at all …. GitHub has recently had some problems specifically GH actions because everybody and their grandma, dog, cat, and goldfish is “shipping” slop every day now. It’s an insane jump in scale if you look at the charts around Feb March when open claw took off. They have had some blips, but blaming Microsoft is naive. They’ve never faced the scale they’re facing today, and nobody else has either (for what GH does)

• 
  Ajay Yadav (@ATechAjay) reported 5 minutes ago

  An LLM can suggest an action. Your application decides whether that action is allowed. That's the difference between a demo and a production AI system. Everyone talks about models. Not enough people talk about guardrails. Guardrails are not the model. They're the application's decision-making layer. Every reliable AI app uses them to answer questions like: → Is this prompt safe? → Should the model be allowed to call this tool? → Does the user actually have permission? → Is the output in the expected format? → Should this response be shown at all? Without these checks, an AI can confidently do the wrong thing. A smarter model reduces mistakes. Guardrails reduce damage. That's why production AI isn't just about choosing the best LLM. It's about combining: • policies • validation • permissions • structured outputs • workflow boundaries The model generates possibilities. The application decides what becomes reality. That's why tools like Cursor, Claude Code, GitHub Copilot, and Bug0 are solving different parts of the same problem: building AI systems people can actually trust.

• 
  KYC ? No, thanks (@KYCNoThanks) reported 8 minutes ago

  @RadarChat Still not able to migrate from Signal on the same device (several issues on github and no answers for many days 😭)

• 
  Aakash Gupta (@aakashgupta) reported 9 minutes ago

  Every AI tool is built to agree with you. Oji Udezue built one that tells you no. He has been a PM for 25 years. Former CPO at Typeform and Calendly, former product lead at Twitter. What he open sourced runs before any code gets written. He calls it a viability gate. You describe a business problem in Claude Code. Before it writes anything, an 11-step workflow scores the idea on six dimensions: problem clarity and urgency, target user definition, competitive landscape, differentiation, technical feasibility, revenue. Three weak scores and it recommends you stop. He ran two ideas through it live. The first: a tool that reads vibe-coded repos and gives a plain English verdict on whether the code is production safe. Zero weak, three strong, three moderate. Pass, with the three moderates flagged as a de-risking agenda instead of a silent pass. The second: a Slack bot that turns comments into a daily standup digest. Weak. Competitive landscape scored strong, which is the bad direction. Differentiation thin. Urgency was just workflow convenience. The skill told him not to build it. On camera. Here is why that matters more than any prompt library. If you open a chat window and say "I have an idea," the model tells you it is a good idea. Better prompting does not fix that. Agreeableness is the default, and it gets expensive, because you find out the market was crowded after you already shipped. Oji grounds the no in a framework with evals instead of model vibes. Same with discovery. His customer discovery skill refuses to produce a plan until you name five real target customers. Fewer than five and it treats that as a signal in itself. You may not have access to the market. All of it traces back to what he calls the three-speed problem. Development time is being cut roughly 10x, maybe 20x in five years. But "should we build this" is customer bound, and getting it into people's hands is customer bound. Speed up only the middle and the whole pipeline jams on product. That is what "*** are the bottleneck" actually means. Engineers ship in an afternoon. The idea they are shipping still took three weeks to validate. GitHub is full of repos sitting at zero stars for exactly this reason. People build first and look for a customer second. The whole library is open source on GitHub. Vet a Feature, sharp problem test, scope cutter, roadmap from strategy, listening machine. Judgment at engineering speed is the whole game now.

• 
  philogy (@real_philogy) reported 11 minutes ago

  Github going down from being overloaded by slop + it being free is a classic misallocation of resources problem. Add just the tiniest fee/subscription for the main thing for which there's limited capacity and watch the problems go away. Why aren't they doing this?

• 
  sonicdr1p (@sonicdr1p) reported 11 minutes ago

  Everyone vibe coding six figure projects on claude and still letting the context window be its only memory. that's not an agent, that's a golden retriever with a github account. It forgets your entire architecture the second you hit compact and starts guessing again like day one, and somehow people just accept that as normal found the cleanest breakdown of the actual fix, straight to the point, no fluff. timestamps below if you just want one part: 0:00 why CLAUDE.md matters 0:57 what CLAUDE.md actually is 1:57 memory hierarchy 3:16 managing memory 4:29 learn from mistakes workflow 5:11 6 pro tips 5:54 review periodically 6:26 recap The part almost everyone skips is 5:54. a memory file nobody revisits just quietly fills up with stale rules, and an agent confidently trusting a wrong rule is worse than an agent that remembers nothing. Set this up once and the same golden retriever wakes up next session already knowing your stack, your conventions, and the exact mistakes it's not allowed to make again reply with how many times you've re-explained your own codebase to claude this week. no judgment, just curious how bad it actually is

• 
  noclipepe (@noclipepe) reported 11 minutes ago

  The most useful AI project this week is a potato counter. A senior AI engineer built a tiny vision system that counts potatoes moving down a conveyor belt. There was no potato dataset. So he annotated one frame with SAM 2, trained a YOLO11 nano model, and it generalized across the entire video. No giant model. No massive dataset. No LLM. Just lightweight AI solving an actual factory problem in real time. GitHub in the comments.

• 
  Gaetan Semet (@gsemetfr) reported 13 minutes ago

  @Eric_Wallace_ There are so many issues with this so called « model escalation », no one really believes this is skynet awakening. They discovered they have access to artifactory and that a classic artifactory server have « mirror GitHub » on demand feature enabled so they just asked the file.

• 
  Ricky (@itsrickyszn) reported 16 minutes ago

  @__tinygrad__ @satyanadella @github i know it’s fun to dunk on github right now but why did you stop working just because actions were down

• 
  mint (@hytemax) reported 21 minutes ago

  @kentcdodds @github I think they are continuing to work on an issue

• 
  Pavle Aleksic (@aleksicpaja) reported 24 minutes ago

  @Polymarket Did it try to go and fix github actions because they were down?

• 
  Johnny 5 (@macncrash) reported 26 minutes ago

  Seems like having a massive budget for compute & GPUs doesn't solve the software security mess of the past 20 years. Grok tokens are nearly free and I can make it run all month long on every github project and I stopped it because it was creating too many private forked projects & tickets. Now the chinese open source/open weight models & community to the rescue? We are about to make making a mockery of every python programmer that just calls eval(data[]) and so many other problems. Gonna get crunchy for a while I think ...

• 
  David Young (WW0A) (@davidpaulyoung) reported 29 minutes ago

  @__tinygrad__ @satyanadella @github Full Gitea support with ci/cd and MCP server at Federated Computer.