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:

• 
  Joe Bond (@josephscottbond) reported 2 minutes ago

  @gregisenberg Hah! 2 anecdotes from my day confirm you’re right. 1. GitHub actions was down today… GitHub seems like they are having an outage every week now. 2. I was trying to figure out how to help a normie coworker contribute to a technical project without making her sign up for GitHub. Made a joke about vibe coding GitHub this weekend.

• 
  LaSean Smith (@lasean) reported 2 minutes ago

  @shezmalik @tbpn Adjusted for inflation Nvidia's paying ~$3.1B than Microsoft paid for GitHub (~$9.8B). More, but may in the margin of error with all of the AI funny money.

• 
  pulpmatrix (@pulpmatrix) reported 3 minutes ago

  Confirmed. Prime Intellect's writeup is public: GPT-5.6 Sol Pro, offline sandbox, hand-rolled a raw Responses API call to the interception proxy and used file_url to make the inference server fetch a GitHub API response for it.

• 
  Rituraj (@RituWithAI) reported 3 minutes ago

  🚨BREAKING: Researchers just proved that every AI agent controlling your Android phone collapses the moment something unexpected happens. A popup appears. The wrong screen loads. A button moves. An app crashes mid-task. All 16 leading Android GUI agents tested. Every single one degraded significantly under these conditions. It's called AnTrap. A benchmark that injects the kind of real-world chaos that exists on every actual Android device — and measures whether AI agents can handle it. Here's what was actually tested. Every Android GUI agent benchmark until now tested agents in perfect conditions. Clean screens. Predictable flows. No interruptions. The agent taps, the expected screen appears, the agent continues. Real phones don't work like that. AnTrap injected four categories of real-world disruption into agent execution: → State anomalies — unexpected popups, wrong screens loading, UI elements missing → Thinking anomalies — the agent's reasoning gets disrupted mid-task → Action anomalies — buttons move, taps misfire, actions produce wrong results → Round anomalies — tasks get stuck in loops, states become deadlocked Then they tested 16 leading models — every major GUI agent available — against all four. Universal vulnerability. Every model. Significant performance degradation. Across the board. Here's the finding that makes this structurally alarming. Some failure modes are fixable. Single-step traps — unexpected popups, misplaced buttons — can be largely resolved by training agents in adversarial environments. The agent learns to handle interruptions. But deep contextual traps — particularly state deadlocks, where the agent gets stuck in a loop it cannot recognize or escape — expose something that training alone cannot fix. An intrinsic reasoning limitation. The agent doesn't know it's stuck. It keeps trying the same approach. It never asks for help. State deadlocks aren't learnable. They're architectural. Here's why this matters for every AI agent controlling your phone right now. ClawGUI. Computer Use. Every phone control agent being deployed to millions of users operates in exactly the environment AnTrap simulates. Real phones. Real apps. Real popups at inopportune moments. Real UI changes between versions. Every agent was benchmarked in ideal conditions. None were tested in real ones. Until now. 11 upvotes on Hugging Face. From Zhejiang University, Yale, and collaborating institutions. Code available on GitHub.

• 
  James Liberty (@JamesLiber54019) reported 4 minutes ago

  The fix was simple: I had to generate a digital VIP pass (a private security token) so GitHub recognizes my app every time it checks in. Works smoothly now without getting locked out. 🦾 (2/2)

• 
  The Daily Viber (@TheDailyViber) reported 6 minutes ago

  Before teams give coding agents more access, they need a boring inventory of the access already hiding in repos. AI TOOLING HAS A SUPPLY CHAIN PROBLEM, AND MOST TEAMS ARE STILL CALLING IT “EXPERIMENTATION”. Wrkr is a posture scanner from Clyra AI for the agentic development mess that now lives inside repos and GitHub orgs. It looks for AI dev tools, coding agents, MCP servers and workflow action paths, then asks the useful question: which of these can write, which controls are visible, and where is the evidence missing? That is the npm audit moment for agents. Instead of “which dependency is vulnerable”, the question becomes “which agent, MCP server or automation path can touch code, credentials, CI or production-adjacent systems?” The comparison is simple. - AI policy doc: sounds mature, finds nothing. - Wrkr scan: maps tools, paths and missing controls. - Vibes: “we have approval gates somewhere”. - Evidence: detected control, declared control, external reference, no visible control or contradictory control. That dry language is the good part. Static scanning cannot honestly promise runtime enforcement. Wrkr does not pretend it can. It can say what it found in a bounded scan, what it did not find, and which path deserves review. That is how security tooling stays useful instead of becoming demo theater. The practical first move is small: scan one repo, generate an Agent Action BOM, inspect the top workflow and action paths, then check credentials, config files and write authority. If a new MCP server or agent workflow appears without an owner, review and minimum permissions, it should light up before it reaches the release path. One honest note: Wrkr is probably overkill for a solo developer with one empty repo and a single Claude Code session. It starts to make sense when an org has multiple repos, GitHub Actions, MCP configs and people adding tools because “it helped on my machine”. Static posture also does not fix the problem by itself. Someone still has to remove the tool, narrow the credential, add an approval gate or document the exception. That is fine. First you need the map. Then you can decide where the borders go.

• 
  Vysakh Sreenivasan (@vysakh0) reported 16 minutes ago

  SPEC driven development. Create spec as github issue. A spec can be executed by any coding agent in anyone's machine.. SPEC > PR

• 
  Harshil Tomar (@Hartdrawss) reported 22 minutes ago

  Even if you're not 17, you can still learn building LLMs from scratch ! Laptop + below Sauce + Hard work = Infinite Money 1. start with @karpathy zero to hero playlist : - micrograd first, 150 lines, you write backprop by hand and never fear autograd again - makemore parts 1-5, you go from bigram counts to a batchnorm MLP and see why initialization decides whether a net trains at all - "let's build GPT", 2h and you have a working decoder-only transformer on tinyshakespeare, self attention, multi-head, residuals, layernorm - the tokenizer video, you build BPE from scratch. this is the one everyone skips and then can't explain why their model is bad at arithmetic that puts you ahead of a lot of people already 2. try to rebuild nanoGPT from an empty file - target is GPT-2 124M, 12 layers, 12 heads, 768 dim, 1024 context - write the dataloader, the model, the training loop and the sampler yourself, no forking - get it training on openwebtext and watch val loss land near 2.85, thats the public benchmark you're checking yourself against - learn what actually breaks: gradient accumulation, mixed precision, cosine LR with warmup, grad clipping at 1.0 3. train a tiny model on one dataset for a specific usecase - 10-50M params on your own data beats another wikitext run for learning - your discord export, every commit message you've written, one author's full catalogue, 20 years of match reports - chinchilla rule of thumb: roughly 20 tokens per parameter, so a 10M model wants ~200M tokens. know when you're data starved - build your own tokenizer for it, a domain vocab of 8k beats a generic 50k on narrow data 4. rent an A100 for $2/hr on vast[.]ai - runpod and lambda price similarly, $30-50 covers a real overnight run - a 124M model on one A100 80GB is roughly a day of training, budget for 3 failed attempts before the good one - do not buy a $3k gpu, thats spending money to avoid spending hours 5. start studying deeper topics : - finetuning: LoRA rank 8-16 on the attention projections, alpha at 2x rank. QLoRA in 4-bit nf4 fits a 7B on a single 24GB card, thats a $0.40/hr machine - inference: KV cache is why memory grows with context, quantize to Q4_K_M for ~4x smaller with small quality loss, use vLLM's paged attention and continuous batching before you blame the model for being slow - evals: write 50 real prompts for your usecase and grade them before you tune anything. no eval set means every finetune feels better and you'll never know 6. study one paper a week and reimplement the smallest part - start order: Attention Is All You Need, GPT-2, Chinchilla, LoRA, RoPE, FlashAttention, DPO - don't reproduce the paper. reproduce one figure, one layer, one loss function 7. build one thing people can use : - a narrow finetune that beats a frontier model at one job, classification, extraction, one specific format - self host it, measure cost per 1k requests, that number is your margin - 10 paying users is the proof, not the github stars 8. post every run in public, failures included : - the loss that diverged at step 4000, the run you killed at 3am, the config that finally worked - share the actual hyperparams and the wandb chart, people follow specifics - compute grants, jobs and co founders find you through this long before a credential does honest math: ~12 months of consistent evenings, about $400 in rented gpu hours.

• 
  Matt (@matthew_hartman) reported 23 minutes ago

  @coleywoleyyy I use github issues for that right now. Not perfect but I like that they are connected to my code and repo.

• 
  Old Jobobo (@OldJobobo) reported 27 minutes ago

  The public-action layer still needs hard limits. Those eight QA agents reportedly found 28 real issues, then filed all 28 on GitHub within roughly 12 seconds. GitHub flagged the activity as spam and banned the Omarchy bot. Valid findings still need de-duplication, rate limits, and approval gates.

• 
  Devon Govett (@devongovett) reported 29 minutes ago

  @rafalfilipek can you open a GitHub issue for this? seems fine to add that.

• 
  Gabriel Rubens (@gabrielrubenss) reported 37 minutes ago

  VPS deploy via GitHub (5/8): for now a blocked deploy simply runs again on a fresh runner, and that rescued five of the next six. It is a workaround though, not a fix. The strange part: it started out of nowhere and I changed nothing in my infra, so I still want the real cause.

• 
  Abhishek (@aacle_) reported 37 minutes ago

  A public GitHub issue → GCP Editor access. That sounds fake, but the attack chain was actually this: Attacker leaves a malicious comment on GitHub. Gemini CLI reads it. The prompt injection gets the agent to execute commands on the CI runner. From there, the attacker finds cloud credentials and impersonates a service account with Editor-level access.

• 
  Kunal (@kunal_twts) reported 41 minutes ago

  It’s all about tokens nowadays. No one cares about APIs anymore. MERN projects… whatever. Documentation… barely. Stack Overflow… slowly disappearing from the workflow. YouTube tutorials… I used to spend hours there. And sometimes I genuinely miss the old internet. I literally had a Blogger page running around 2016–17 where I used to write posts about random stuff like how to download and run compressed GTA V, tweaks, fixes, tutorials and whatever I was obsessed with at the time. And the crazy part? That little blog hit 97,000 views at its peak. 97 ******* THOUSAND. I mean… damn. I was just some kid writing tutorials on Blogger, probably copying half the knowledge from somewhere else, figuring things out as I went, and somehow thousands of people were landing on my page because they had the exact same problem I was trying to solve. Back then, the internet felt so much more… alive. I learned face recognition in 2017 by basically following YouTube tutorials line by line. Copy the code. Run it. Error. Google it. Find a Stack Overflow answer from 2014. Copy that. Another error. Back to YouTube. And when it finally worked, it felt like I had discovered some forbidden technology. Animation felt hard. Photoshop felt hard. After Effects felt hard. Even making some stupid PicArts edit on a phone felt like a skill. There were Opera Internet hacks, random blogs, shady forums, GitHub repos, YouTube tutorials with 300k views and a guy explaining everything with Windows Movie Maker-level editing. There was friction everywhere. But that friction made you learn. Now? I have an idea and AI can build 80% of it before I’ve even opened the documentation. Which is absolutely ******* insane. But the new friction is: tokens. And the constant anxiety of burning them. You start building something. “Fix this.” Tokens gone. “Actually make it responsive.” More tokens gone. “Refactor the whole thing.” More tokens. “Wait, I didn’t mean that.” 💀 And then you’re sitting there thinking: “How much do I have left before the next wave resets?” We went from: “I need to figure out how to build this.” to: “I need to make the most of these tokens before they run out.” Maybe that’s progress. It definitely is. But sometimes I miss being that kid with a Blogger page getting 97,000 views because I wrote a ****** tutorial about GTA V. I miss opening YouTube to learn something. I miss breaking code for hours. I miss finding some random Stack Overflow answer that saved my entire project. I miss the feeling of actually earning the result. The internet has become insanely powerful. But damn… the old internet was fun.

• 
  Jarrod Watts (@jarrodwatts) reported 43 minutes ago

  @colemurray @tryreplicas Sick, I setup 1 & 3 so far for an app I'm building. It's so cool seeing PRs raised from sentry issues I'm going to do a github one today to respond to issues & PRs to my claude HUD repo