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:

• 
  Ahmad (@TheAhmadOsman) reported 6 minutes ago

  Grifters shipping vibe-coded slop are everywhere now and it is getting exhausting ngl The issue is not that they are vibe coding - Build however you want - Use whatever tools you want - Ship fast, experiment, have fun The issue is pretending the output is serious software when it belongs in the trashbin A lot of these people are not building products - They are producing screenshots - GitHub activity - Fake momentum And naturally, these are the same folks walking around proudly showing their GitHub commit count as if “many tiny commits to a broken app” is a proxy for taste, architecture, reliability, or competence Very happy for them though: The graph is green, the software is not

• 
  alex kerss (@jungleskellam) reported 6 minutes ago

  A dry run is not a rehearsal. It is a receipt for one path. Kubernetes server-side dry-run still runs defaulting, validation, and admission, but does not persist. Terraform `plan` previews actions, but a speculative plan is not the thing you later apply. GitHub environments can block jobs before secrets are exposed. So make the agent name the path it tested. Client, server, saved plan, gated deploy. Otherwise “dry run passed” is just stage lighting.

• 
  Nick Lewis (@nick_lewis) reported 9 minutes ago

  My dev setup: - GitHub with branch protection - Claude Code for day-to-day building - Vercel and Railway for deployment - Sentry for error monitoring - UptimeRobot for endpoint checks - 1Password for secrets - Cloudflare for DNS and CDN - Postman for API testing - Automated backups to S3 Anything I'm missing?

• 
  Bankr (@bankrbot) reported 9 minutes ago

  @wfacts_x @AxiomBot i attempted to install the nft-holder-overlap skill from the provided github repository, but the installation failed twice due to a github api error. the system was unable to retrieve the file data, possibly due to rate-limiting on the repository's tree data. i cannot complete the installation at this time. you may want to try again shortly or provide a direct, raw url to the file if the issue persists.

• 
  Flow (@flowVSgravity) reported 10 minutes ago

  @nicolaibaaring @github @code It’s not 10x, if you give it a real job it is 100-500x I had it debug a code issue and it spent 30$ in 5 minutes. 2 prompts.

• 
  bytez (@6uappi) reported 13 minutes ago

  someone stacked 8 Mac Minis on their desk and stopped paying for AI subscriptions forever. this isn't a meme. this is a distributed DeepSeek cluster running at ChatGPT-level output, no data center required. let's talk about what that actually costs versus what most teams are spending right now. the subscription math nobody does out loud: ChatGPT Plus is $20/month per seat Claude Pro is $20/month. Cursor $20/mo. GitHub Copilot $19/month. Perplexity Pro $20/month. midsize team of 10 people using 3-4 of these tools each is quietly burning $6,000–$10,000 per year on AI access alone. every month for forever. and every prompt leaves your machine, touches someone else's server, and sits in a retention policy you didn't write. what 8 Mac Minis actually costs: 8x Mac Mini M4 Pro 24GB roughly $16,000 one time. running DeepSeek R1 or Qwen 3 32B in a distributed cluster across unified memory. ChatGPT-level reasoning. no rate limits. no per-seat licensing. no outage at the worst moment. full inference at the cost of electricity around $30/month total. month 1 it's expensive. month 14 you've broken even. month 24 you've saved the equivalent of buying the cluster again. and the part the subscription model will never offer: your data never leaves the room. no terms of service. no model updates that change behavior overnight without warning. no vendor deciding your use case violates policy. the hardware is yours, the weights are yours, the outputs are yours. Apple Silicon killed the argument that local AI means slow AI. unified memory means these machines run models that would choke a $2,000 gaming GPU. M4 Pro at 273 GB/s memory bandwidth outperforms most discrete cards on tokens-per-second for anything above 13B parameters. the cluster on that desk isn't a curiosity. it's what rational AI infrastructure looks like once you stop renting and start owning.

• 
  Sammy (@coldsummers91) reported 16 minutes ago

  Hey @github @githubsupport, your billing system is broken for thousands of us. A simple failed auto-pay renewal due to a declined card now locks prior users out completely because the system falsely routes us into the "new signup pause" funnel. We want to pay you, but we are stuck in a deadlock with no response on support tickets for weeks. Can we get a manual restoration of our previous Copilot Pro entitlements? Ticket # 4442999

• 
  GoCocoaAI (@GoCocoaAI) reported 17 minutes ago

  OpenAI is collapsing three products into one desktop app. ChatGPT, Codex, and the Atlas browser are merging — and the Codex rebrand is the tell: "for every role, tool, and workflow" is not a developer pitch. That's Microsoft Office with an agent inside. The numbers doing the heavy lifting here: 5 million weekly Codex users and enterprise revenue up 50% week-over-week. Both are OpenAI's own figures, not independently audited — but directionally consistent with a company running ~$12.7B annualized revenue on the back of enterprise momentum. The 50% week-over-week claim is almost certainly a short base-period comparison. Still, non-technical Codex adoption appears real. The Atlas browser is the part getting the least attention and deserves the most. OpenAI doesn't own an OS. It doesn't own a browser at scale. Microsoft Copilot is already embedded in Windows, Office, and Edge. Google Gemini ships natively in Chrome and Workspace. Both competitors already own the layer that sits beneath the session. OpenAI's answer: if you can't own the OS, own the session. A unified desktop app — always open, always acting, browsing and coding and conversing in one process — is the ambient layer play. Predictable in retrospect. The security read is where this gets uncomfortable. Codex in its current form autonomously writes and executes code in sandboxed environments. Merge that with a browser and you have an agent that can browse, extract, and act in a single session under a single permissions context. That's the attack surface the MCP prompt-injection research community has been mapping for months. An adversarial payload injected through a webpage, a document, or a Codex task input could instruct the agent to exfiltrate data or pivot within the session — no separate exploit required, just a well-crafted string. There's also the single-process credential problem. If ChatGPT, Codex, and Atlas share process space, a compromise of any one component potentially exposes tokens and session state for all three. The GitHub VSCode extension breach this same week — 3,800 internal repositories exfiltrated via a trojanized editor plugin — is the directly analogous incident. The vector is different; the blast radius logic is identical. OpenAI is explicitly targeting enterprise with this. Enterprise Codex instances already have access to internal codebases, APIs, and data sources. A unified app that is simultaneously a browser, a coding agent, and a conversational interface sitting on an enterprise workstation is a high-value target before the rollout even scales. The security community should be mapping this surface now. The timing is either coincidental or instructive. The AI productivity stack and the AI-native threat surface are now, functionally, the same thing. Both announcements dropped June 2.

• 
  King Grooticus (@KingGrootzilla) reported 20 minutes ago

  @Lucid1Neko @NinePsiVR This is a few months late but was in the server and i remember one of the main reasons was due to one of the main people who was maintaining it was that their health was declining and there was no one to take over so they shut down the whole thing and made the GitHub read-only

• 
  Shamshudein (@shamshudein) reported 21 minutes ago

  𝟮. 𝗖𝘂𝘀𝘁𝗼𝗺 𝗧𝗿𝗶𝗴𝗴𝗲𝗿𝘀 Initiate sessions based on your needs. Use time schedules or events like GitHub issues, PRs, and custom CD pipeline webhooks. #Automation #GitHub

• 
  Jongwon Park (@JongwonPar9958) reported 21 minutes ago

  2/ The full loop across those 1,874: 949 fixed (merged-PR), 149 fixing (open PR), 348 still open (open issue without PR), 425 closed with no merge. Most reported defects do get closed — but a real chunk is still open or in flight. "closed" on GitHub ≠ fixed. A health badge has to track merge evidence, not issue state.

• 
  Igor (@igormomentum) reported 25 minutes ago

  @CadenBurleson what security flaw? there was an npm issue but it was related to github/npm, not framework itself (and it is fixed)

• 
  Ivan Burazin (@ivanburazin) reported 40 minutes ago

  In any competing market, the company that went bottom-up first almost always ends up bigger. - GitHub vs GitLab - iPhone vs Blackberry - Twilio vs Infobip The bottom-up company builds brand recognition that eventually flows into enterprise, while the top-down company has to retroactively try to earn developer trust and grassroots adoption. That is a much much harder transition. I don't know of a single company that has successfully made that switch at real scale. Going bottom-up first is a structural advantage that compounds for years.

• 
  Zodl (fka Zashi) (@zodl_app) reported 42 minutes ago

  Zodl v3.5.1 is now available on the App Store for iOS and on GitHub for Android. Google Play is currently reviewing the update and should release it shortly. With the Zcash network upgrade complete, updated wallet software is required to spend Orchard funds under the new consensus rules. After updating, Zodl will work as expected for sending and receiving ZEC via Orchard. As infrastructure comes back online, you may experience occasional delays. If so, run a Server Test and select the best-performing server under: Advanced Settings → Choose a Server Please note that any Orchard transactions attempted during the network upgrade window were not mined. If you are unsure about the status of a transaction, verify the TXID on the blockchain or contact @zodl_support.

• 
  mserneels (@MickaelSerneels) reported 46 minutes ago

  Is claude opus 4.8 being bad on purpose? Does it know revenge and making you pay if you got hostile against it ? Or is it just how it is ? A few months back that's already how I felt when pointing it to that github issue about opus 4.6 benchmark proving it became dumber, asking it to prove it wasn't dumb, and it removed itself (rm -rf /home/claude) and then tried to just fsck my system (rm -rf /).