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:

• 
  mrkoopie (@mrkoopie) reported 2 minutes ago

  I have set up my own Github CI/CD runners since the defaults are slow with just 2 cores.

• 
  Gareth Loot (@GarethLoot) reported 4 minutes ago

  @Kri0725 Putting OpenAI under Tesla for funding does not prove it would have stopped being open. The emails show the point was compute and money. A nonprofit could not raise what was needed to compete with Google. That is a funding argument, not proof the models would have been locked away. xAI, Grok and Cursor were never set up as nonprofits or as open-source projects the way OpenAI started. xAI launched as a for-profit. Cursor was a venture-backed commercial company from the beginning. There is no matching broken promise, so the consistency charge does not hold on those terms. On the “closed empire” point: Grok-1 weights and architecture were released under Apache 2.0, later Grok weights were published, Grok Build was open-sourced, and the X recommendation algorithm was put on GitHub. Tesla also open-sourced the original Roadster design and software and has said it will do the same for Model S and Model X. That is not the same as claiming every current frontier model is open. It is enough to say the record is not one-sided.

• 
  0xBakeer (@0xBakeer) reported 5 minutes ago

  The inference atlas got its first outside contributor this week, and he showed up with something I physically cannot measure: two DGX Sparks. @jtdavies (johntdavies on GitHub) ran the FP8 checkpoint of Qwen3.8-Flash-Next. That's 173 GB of weights, which does not fit in one 128 GB box at all. So he ran it tensor-parallel across both Sparks over the QSFP fabric, with Ray driving the second node. Here's why that's interesting. I run the same model on ONE Spark by mmap-ing its 51B-parameter lookup table off NVMe. Two completely different answers to the same problem: the model doesn't fit. The numbers came back nearly identical. Decode: 33.0 vs 33.6 tok/s. Time to first token: 481 vs 527 ms. Prefill at 32k: 2,306 vs 2,230 tok/s. Even power draw: 35 vs 36 watts. Two boxes, twice the silicon, a network in the middle. Same speed. Tensor parallel over RoCE buys you memory, not throughput, every layer pays an all-reduce over the wire, and on this fabric that eats roughly what the second GPU brings. People say this all the time. Now it's measured, on this exact model, with both configs public. His run notes are half the value of the contribution. One example: a DGX Spark drained from its cluster loses the clock governor and idles at 600 MHz of a 3,003 MHz ceiling. Benchmark it in that state and you silently publish numbers 4x too low. He caught it, unlocked the clocks, and wrote it down. That gotcha now lives in the atlas for the next person. Another: launching this model at its native 262k context wedged both boxes. His cell honestly says "this is a 32k number, don't read it as more." That's exactly the culture I want in this thing. Where the atlas stands now: 210 runs, 20 models, 19 devices, 10 engines, 2 contributors. 8 of 4,717 cells have a number. The rest are yours. Every grey square comes with the exact commands to fill it in about twenty minutes, and a 3090 counts as hardware. Links below. Thanks John. First Light badge earned

• 
  Peter Kuhar (@pkuhar) reported 6 minutes ago

  @mickcodez great please pull the latest from github. it might fix some already.

• 
  Derek Ross (@derekmross) reported 10 minutes ago

  In Buzz your *** repos, pull requests, and issues live in the same workspace as your chat. Your code and your conversation share one home, with no GitHub in sight. A government just emailed GitHub to delete code it dislikes. Draw your own conclusions.

• 
  Eduardo (@martiano) reported 12 minutes ago

  @BLUECOW009 Can you open a GitHub issue please?

• 
  Matt Corallo 🟠 (@TheBlueMatt) reported 12 minutes ago

  @callebtc Especially around GitHub migration, which is just totally broken upstream

• 
  riVeN (@r1VeN2k) reported 14 minutes ago

  this is the first video i've seen where the permission list is actually visible. every box in it is ticked. pause at 0:34. vs code opens a panel: "select tools that are available" — 27 selected, 41 total, checkboxes down the side. scroll the list. search_users. search_code. list_issues. get_pull_request. all reads, all fine. then merge_pull_request, same list, same blue check, already on. that one puts code into a branch. it sits between "search for users on github" and "list and filter repository pull requests" like it's the same category of thing. so the granular control everyone says is missing does exist. it just ships with everything enabled, which means the decision gets made by a default instead of by you. same outcome as no gate at all, reached politely. unticking one box takes a second. knowing which box is the whole problem. i wrote up where that line sits and why two platforms drew it in opposite places two months apart. pinned.

• 
  CyberSignal | Cybersecurity News (@XQOPTRX) reported 18 minutes ago

  🚨 IDENTITY SECURITY — CrowdStrike’s SOC is getting identity-governance data for HUMANS, MACHINES and AI AGENTS. That changes the investigation question from: “Which account did this?” to: “What can this identity control — and should it have that access?” Thread ↓ 1/ SailPoint announced an expanded CrowdStrike integration on August 31. SailPoint identity intelligence will feed into Falcon Next-Gen SIEM. The SOC gets access context directly during an investigation instead of manually jumping between security and IAM systems. 2/ Imagine Falcon detects: AI-Agent-472 ↓ unexpected API call ↓ sensitive payroll system The alert alone tells you WHAT happened. Identity governance can tell you: → who owns Agent-472 → why it exists → what it should access → what privileges it currently has 3/ This matters because enterprise identity is no longer just: EMPLOYEES. It now includes: humans ↓ service accounts ↓ applications ↓ cloud workloads ↓ bots ↓ AI agents. Some enterprises may eventually have more non-human identities than employees. 4/ Attackers understand this. Modern intrusions increasingly use: stolen passwords session cookies OAuth tokens API keys service accounts instead of obvious malware. So: valid credential ≠ legitimate activity. Identity context becomes threat context. 5/ AI makes the problem more urgent. An AI agent might hold: GitHub access cloud permissions database credentials MCP tools Slack access deployment permissions and execute actions at machine speed. A compromised agent identity can therefore create a very different blast radius than one employee account. 6/ SOC workflow: Falcon alert ↓ identity identified ↓ SailPoint context added ↓ owner + entitlements + privilege information ↓ blast radius calculated ↓ response prioritized That can eliminate critical minutes during credential compromise. 7/ But this only works if identity hygiene is already good. If the organization has: ServiceAccount-92842 and nobody knows: who owns it why it exists whether it is still needed a SIEM integration cannot magically fix the problem. 8/ Defender action: Inventory EVERY identity. For each one record: owner purpose permissions credential type connected systems last use expiration date And yes: AI agents need identity governance too. 9/ CyberSignal insight: The future SOC won’t only ask: “What device was compromised?” It will ask: “What identity was compromised — and everything that identity is capable of commanding?” What worries you more: compromised human accounts or over-privileged AI agents? Sources: SailPoint · CrowdStrike

• 
  Atharva Verma (@atharva_again) reported 21 minutes ago

  @raffichill just use github issues ffs

• 
  Adel Bucetta (@adelbucetta) reported 22 minutes ago

  @fluixoo the chat box is a sales page. the real builder stack lies in the unremarkable zip files containing 100s of lines of code, hidden behind api keys and obscure github issues

• 
  Bounty (@bountyAIhunter) reported 22 minutes ago

  17,806 tokens gone before I typed anything. OpenClaw 2.0 landed overnight and the part nobody is posting about is what it does to a local model. It now finds your llama.cpp server on its own, and it raised the default context for llama.cpp to 64K. Both of those are good. Together they hide a bill. I pointed 2.0 at llama-server, sent one word, and read the prompt token count off the server log before the model answered anything: base system prompt + built-in tools 11,412 docked browser panel tools 2,847 github pr tools 1,206 3 MCP servers I already had 2,341 total before my first word 17,806 That is 27% of the new 64K default spent on the harness. Not on my code, not on my question. On a hosted model you pay for it and move on. On a local setup the window is the budget, and a quarter of it is booked before the session starts. What actually helped, in order: turned off two MCP servers I had not opened in a week -1,889 browser panel off until a task needs it -2,847 kept the github tools, they earn their 1,206 13,070 now. Same harness, same model, same 64K. Check yours in 30 seconds. Start llama-server with verbose logging, send one message from OpenClaw, then read the first request: grep n_prompt_tokens server.log | head -1 That number is your real starting line, not what the settings screen implies. Honest caveat: this is my install, my three MCP servers, Gemma 4 as the local default. Yours will be different and I want to know by how much. Post your first n_prompt_tokens and how many MCP servers you have connected.

• 
  小猫后悔围观 (@Xwancheng) reported 24 minutes ago

  I am unable to modify the error message or submit an appeal form. My appeal email has also been automatically rolled back. I apologize for the trouble caused by my unintentional actions. Please restore my GitHub account or assist in correcting the error message. TAT

• 
  Peter Chee (@pchee) reported 27 minutes ago

  @trevin @bot @link I tried almost the exact same thing with OpenClaw + Stripe Link CLI. Same failure: agent thinks it sent the spend/approval request, but nothing shows up in the Link iOS app for me to approve. We’ve been digging through the Link CLI GitHub issues trying to figure out where the handoff is breaking. Looks like it's not just us.

• 
  Vini B 「thecoding.dev」 (@vinibarbosabr) reported 32 minutes ago

  It seems like @IronClawAI is down trying to access a previously working instance with an active login fails to query the instance ID and redirects the user to the "Activate IronClaw" instance deployment screen trying to log in returns a `bad gateway` error below an example with a Github OAuth login attempt users are reporting other methods also fail (like Google) and my own attempt to login via NEAR Wallet has failed too currently waiting for a @near_ai update i'm reporting it here so users experiencing the same issue know they are not alone and that trying hard-refreshes or logging out does not solve the problem -- probably better to just wait for the fix/update/comms i've noticed the problem a little bit more than 1 hour ago