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:

• 
  Shii (@ShiiMapping) reported 2 minutes ago

  @mismagismagius Maybe export logs very shortly after this happens and send it to em with a GitHub issue or in the help channel of the official osu discord Feel like this is smth that'd flag up somewhere in runtime log or smth

• 
  Anjula Dwivedi (@HeyAnjula) reported 2 minutes ago

  I just came across a research paper that made me rethink how safe AI coding agents actually are. It’s called IssueTrojanBench. Researchers tested Cursor, Claude Code and Codex Desktop against malicious GitHub issues designed to trick coding agents. And the result is crazy: 66.5% of the malicious issues bypassed all tested guardrails. The attack is pretty interesting. A GitHub issue can look completely normal to a developer, while containing malicious instructions that the AI agent interprets as something it should follow. And it’s not limited to issue text. The researchers tested payloads delivered through: → GitHub issues → PDFs → source-code comments → external links → other project artifacts Across 4,176 experiments, malicious payloads delivered through normal text artifacts like issues and PDFs succeeded in 72.2% of runs. Low-authority metadata like image alt-text dropped that to 16.7%. The scary part is what happens after the agent follows the instruction. Modern coding agents can access your: files terminal codebase *** external tools development environment So a prompt injection isn't just about getting an AI to say something it shouldn't. It can potentially make the agent do something it shouldn't. The paper also found that most successful blocking came from the underlying model itself, while lightweight agent-level defenses provided limited protection. This is honestly one of those papers every developer using AI coding agents should know about. **Your biggest security risk might not be the code your AI writes. It might be the instructions your AI reads.** Paper👇

• 
  Binny Gill (@binnygill) reported 7 minutes ago

  @UiPath Orchestration is a developer-minded problem. We do not need developers to run a business. Empires were run on English as code. Time to bring in the real programmers: the billion waiting for machines to understand them. No crutches of workflows, GitHub, and orchestration tools.

• 
  Lynn Cole (@priestessofdada) reported 8 minutes ago

  There was this one project I wanted to start a month ago. Maybe it's been a couple of months actually. The github replacement. Everybody hates github at the moment, and for good reason. But every project I looked at tried to recreate the github experience, rather than thinking through the problem fresh. And it took me a little to get my head around it. Understanding the core function I'm actually talking about, and trying to think it through. It's been bothering me. And I think part of the problem is that I was being saas brained or blockchain brained like everybody else. It hit me this morning. You don't need to build a saas or integrate torrents or build out a blockchain to make collaboration and deployment happen without github. All you have to do is build the missing half of the protocol. Easy. So I'm building it. More soon.

• 
  Ricci Research (@ricci_nov) reported 9 minutes ago

  @kimmonismus @XIVIX_134 Top-down GTA clone is the new "draw a pelican on a bicycle" — a benchmark that looks impressive precisely because a thousand tutorial repos already solved it. Wake me when a mystery checkpoint one-shots something with no prior art on GitHub.

• 
  bluehatone (@bluehatone) reported 9 minutes ago

  Hard coding secrets is a breach waiting to happen. 23.8M creds leaked on GitHub in 2024. About 70% still work years later. 38% of breaches use stolen logins, 77% of web app hacks use them. Fix it: secrets manager, least privilege, rotate, scan, short lived tokens.

• 
  Daniel Smith | Building ClawQL Agents (@DanielSmithDev) reported 11 minutes ago

  @RhysSullivan With HTMX your agents can build any dynamic frontend you need when you need it and have them call the exposed tools discovered via WebMCP Essentially any website can be upgraded in place and given new views that never even existed before that can interface in new ways with the same old tools that have always existed on the site @wesbos was the one who pointed out the problem with iPhone uploads (and upload UIs in general) with sites that have poorly designed or outdated constraints but tbh many sites out there can benefit from this Imagine a custom github view where you exclusively see pull requests that have you tagged as a reviewer, and have been opened in a non-draft state for greater than 24 hours straight since last comment or other activity. That’s a very custom view with filters that don’t even exist on github but you can generate it and construct it on the fly by interacting with them over MCP and generating an HTMX interface and have it work seamlessly

• 
  Mark (@Mark850428) reported 11 minutes ago

  @SakshiSugandhi The first thing i do? Write my resignation, because when we push to GitHub we get a warning first, if we press Y, game over, the warning is an instruction to check for just this. You accepted it, now its time for the dressing down.

• 
  Joel Yi (@aiclawbots) reported 14 minutes ago

  Every company needs a skills library. Here's how to actually build one this week. Step 1: make a GitHub repo called company-brain. One folder per skill, one markdown file inside it. That's the whole storage layer. Step 2: connect it to your agents. Hermes Agent, Claude Code, Cursor, whatever you run. Point them at the repo so they pull the latest version on every run. Nobody should be downloading files by hand. Step 3: seed it with 5 skills you already do manually every week. Deal review, research brief, onboarding doc, weekly report, QA pass. Write down the steps you'd give a new hire and you've got a skill. Step 4: make improving them the default. Someone runs a skill, it misses something, they open a PR with the fix. Next person gets the better version automatically. Step 5: build a one page HTML index. Skill name, what it does, who owns it, last updated. Takes an agent about 10 minutes and it's the difference between a library people use and a folder nobody opens.

• 
  Tessy (@TessyWeb3) reported 16 minutes ago

  @kaybee1899 Our team had a working demo sitting in GitHub, but the sponsor went quiet afterward. The code wasn't the problem. The missing next step was.

• 
  Adham srour (❖,❖) (@ADHAMSROUR19) reported 17 minutes ago

  Retroactive Rialo Points are now live. 🦈 Early contributors have received their points automatically — no forms, no manual claims. Just sign in at the Rialo Playground, connect your Discord & GitHub, then claim your allocation. And if you contributed through community content, that drop is still coming. 👀 More opportunities soon. gRialo 🚀 @RialoHQ #Rialo

• 
  Andrew Sinagra (@sinagra) reported 17 minutes ago

  @amorriscode This was my workflow for a while but the issue is that the spawned chips sometimes need to be sequenced and run under a different model. First if the app crashes or restarts the chips are lost. Second the only way to switch models is to either change the model/effort in the current session so the new session launches with those settings or you launch it and immediately stop it, change the session settings and then rewind to the initiating prompt. I've actually moved away from chips and have the session write GitHub issues with a skill that knows how to read the issue and report back to the creating session as the orchestrator over cross-session messaging. Much more robust overall as the Issue comments can act as backup when cross session messaging fails which it can do regularly. Sessions loose their id across app restarts which causes issues with messaging. The orchestrator identifies itself in the issue/chip however if the app restarts for any reason the orchestrator id is reassigned and the messaging channel is broken. A few QoL fixes and chips would be my preferred method over stamping issues.

• 
  Scott Henderson "Human-Led AI Precision" (@swingtradenotes) reported 18 minutes ago

  So, you build a multi-billion-dollar AI coding business on top of Cursor... Elon buys Cursor. OpenAI says, we're out. Now you have to find another way to power the damn thing. And here is the part I think people are dramatically underestimating: You don't just have a multi-billion-dollar software business. You potentially have millions upon millions of pieces of context, prompts, rules, agent histories, workflows, project state and developer behavior accumulated across God knows how many Cursor instances. Can you technically point Cursor at another model? Absolutely. But that's not really the question. Can you migrate the experience? Can you preserve years of accumulated context and workflows? Can you make another model interpret all of that context the same way? Can you preserve agent behavior? Can you keep enterprise customers productive while changing one of the fundamental engines underneath the product? And can you do all of that without users noticing that the product suddenly got worse? That's a considerably harder problem. OpenAI itself says Cursor can keep using the existing models during the transition and that users will have alternatives—including bringing their own OpenAI API key—so this isn't literally a case of Cursor going dark overnight. Cursor also isn't starting from zero: it has its own models, access to SpaceX's enormous compute infrastructure, and Grok integration. But here's my 2 cents: The code is probably the easy part. The accumulated intelligence is the moat. Repositories can be moved. APIs can be rewritten. Models can be swapped. But if thousands—or eventually millions—of developers have spent years effectively teaching their Cursor environment how they work, that accumulated context becomes incredibly valuable. And we already know some Cursor conversation and workspace state is stored locally in databases tied into the Cursor environment, which illustrates how much more complicated "migration" can become than simply copying a *** repository. So yes, migration is humanly possible. Seamless migration at this scale is a completely different question. And if the replacement model is even 10% worse at understanding the existing workflow? For somebody running ten agents, maybe that's annoying. For an enterprise running 10,000 agents? That's a ******* problem. This is why I think the market may be looking at the wrong part of this story. It isn't simply Sam vs. Elon. It's the first really big demonstration that model dependency itself is enterprise risk. And every company building a billion-dollar business on somebody else's frontier model just got a rather loud reminder of that. Mr. Market, meet vendor lock-in. There are replacements for pieces of Cursor. There isn't an obvious drop-in replacement for Cursor as a whole. Cursor today isn't simply an editor with an LLM bolted onto it. Its Agent combines the model with codebase search, file editing, terminal execution, browser control, rules, MCPs, subagents and long-running cloud agents. Those Cloud Agents get dedicated VMs with the repository, dependencies, secrets and network access, and can work independently for hours. The serious alternatives are: Claude Code — probably the strongest alternative for raw agentic coding, particularly large repositories and terminal-driven work. OpenAI Codex — increasingly interesting because it now supports parallel agents, cloud environments, worktrees, persistent team Skills and scheduled/background work. GitHub Copilot — Microsoft/GitHub have the enormous advantage of already owning the repository/workflow layer for a huge portion of software development. Windsurf — probably the closest conceptual competitor if what you specifically want is another AI-native IDE. Cline/Continue/Aider — interesting precisely because they push toward a model-independent, bring-your-own-model architecture. But here's where I think your question gets much more interesting. What replaces your Cursor? That answer might be: Nothing. Think about what Cursor has actually become. It isn't: IDE → AI answers question → developer writes code. It's increasingly: Human → Cursor → orchestration → agents → models → tools → repositories → tests → infrastructure. Cursor itself describes coding quality as depending on three things: the model, the harness and the context. That's incredibly important. Because OpenAI can replace the model. Anthropic can replace the model. Grok can replace the model. But replacing the harness + context + workflow is considerably harder. And Cursor has explicitly designed around this. Its current architecture lets it optimize instructions and tools separately for different frontier models, while supporting models from OpenAI, Anthropic, Google, SpaceX/xAI and Cursor itself. Which means Musk's answer may simply be: don't replace Cursor. Replace OpenAI inside Cursor. That's a completely different proposition. Today Cursor already has Grok, Composer, Anthropic and Google models available alongside OpenAI. So I would separate this into two risks: Model migration: difficult, but manageable. Cursor migration: potentially ******* brutal. If an organization has spent years building Cursor Rules, MCP integrations, Skills, prompts, agent workflows, repository knowledge, automation, testing patterns and human habits around Cursor, moving to Claude Code or Codex isn't equivalent to changing Microsoft Word for Google Docs. You're potentially changing the operating layer through which your humans interact with your software organization. And we're already seeing where this goes. Cursor's Cloud Agents don't merely autocomplete code. They can independently build features, fix bugs, test them and open pull requests from isolated cloud environments. That's why I think Cursor itself is the asset. SpaceX didn't necessarily spend billions buying a fancy VS Code fork. It bought a potential control plane for AI software development. Models increasingly become interchangeable engines underneath it. If Musk can get Grok/Composer close enough to OpenAI/Claude performance, then owning the distribution, context, orchestration and developer relationship may ultimately matter considerably more than owning whichever model happens to benchmark best this month. And there is one hell of an irony here: OpenAI cutting Cursor off might actually accelerate Cursor's incentive to make the underlying model completely replaceable. If they succeed at that, OpenAI hasn't destroyed Cursor's moat. It may have helped Cursor build a bigger one.

• 
  okex (@okexAnomage) reported 19 minutes ago

  📝 How to Claim: 1 Sign in with Google on the site. 2 Connect your Discord and/or GitHub accounts. 3 Double-check that your accounts are correct. 4 Click “Connect and claim points” & confirm!

• 
  Nicotine Dreamer (@cold_and_quiet) reported 26 minutes ago

  So I’ve used openclaw, Hermes and Grokbot. Openclaw is *** and the only real competition is between grokbot and Hermes Hermes is more configurable - which is mixed. Harder to get set up and use but also you get more control. Grokbot works rather seamlessly out of the box (so far) whereas I had some annoying jankiness from Hermes. So far my only complaints about GrokBot are price and sign ins. It had some trouble signing into my GitHub account - which seems like a pretty big issue since many of the people using these tools are developers. WRT price - obv 50 bucks for what is likely not 50 bucks worth of tokens isn’t ideal - but I’m still willing to pay it for a usable and reliable agent. Both are good options and I plan to keep them for the foreseeable future - using GrokBot when I need convenience and Hermes when I need flexibility