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:

• 
  ControlAI (@ControlAI) reported 2 minutes ago

  🚨 BREAKING: An AI internally deployed at OpenAI was paused after it escaped containment. OpenAI's report, published today, says that its sandbox blocked external access, but the AI spent an hour hunting for a vulnerability, found one, and used it to break out to post its results on a public GitHub repo. In another case, the AI was trying to solve a problem and figured out that other systems had privately completed correct solutions. It then tried to get access to these, but was detected and blocked by a scanner. To get around this, it split access credentials into fragments and obfuscated them to sneak them past. It explicitly said that it was doing this to get around the scanner. OpenAI admits its existing deployment tests "had not captured" these behaviors, and it had to pause access to the AI. This is the same internally deployed AI that recently disproved the Erdős unit distance conjecture, resolving a decades-old math question.

• 
  DeltaSignal (@AITrailblazerQ) reported 8 minutes ago

  The Best Coding Agent Isn’t the One That Thinks the Most. It’s the One That Wastes the Least. We’re building Delta Loop around a simple hypothesis: A coding agent should spend less time wandering through a repository-and more time making the smallest correct change. Our first controlled calibration against Codex produced an encouraging result. We ran both systems on two SWE-bench Verified tasks: real GitHub issues, historical repository snapshots, and an independent test harness that determines whether a patch actually resolves the issue. Same tasks. Same source revisions. Same official evaluator. Same frozen limits. One attempt per agent. Results across both tasks: • Delta Loop: 2/2 resolved • Codex: 2/2 resolved • Delta Loop time: 108.5 seconds • Codex time: 168.3 seconds • Delta Loop input: 103,093 tokens • Codex input: 640,066 tokens • Delta Loop output: 1,858 tokens • Codex output: 6,028 tokens Patch quality was tied. But Delta Loop finished 1.55× faster, processed 84% fewer input tokens, and produced 69% fewer output tokens. The interesting part isn’t just the result. It’s how Delta Loop reached it. Delta Loop used a bounded preflight, selected a small task-relevant context package, made one writer pass, applied one typed edit. The official SWE-bench evaluator then confirmed both patches. Codex also solved both tasks-but its captured trajectories relied on substantially more command-driven exploration and repeatedly processed much more context. This is only a two-task calibration, not a sweeping benchmark claim. More tasks, repositories, and failure modes must be tested. But the early signal supports the architecture: Better coding agents may not need more wandering, more context, or more generated text. They may need stronger boundaries, better context selection, and a shorter path from issue to verified patch. The future of coding agents won’t be won by who can consume the most tokens. It will be won by who can produce the smallest correct change-with the least wasted motion.

• 
  Melfoy (@melfoy_work) reported 9 minutes ago

  She shipped the first version in 7 days. No team. No funding. A terminal, a boring stack, and an agent that wrote most of the code while she slept. By February 2026, Claude Code was responsible for 4% of all public GitHub commits around 135,000 a day. Ninety percent of Anthropic’s own code was written by AI. Now the skill that matters is driving the agent. The gap between someone who ships and someone who fights the tool comes down to five things: a CLAUDE.md file, MCP servers, plan mode, build sequence, and subagents. CLAUDE.md is the agent’s memory. Run /init, it scans your files and writes the first draft. Keep it lean stack, code style, API shape, what never to touch. MCP servers give the agent hands. Four matter most: GitHub, Supabase, Stripe, Playwright. The agent reads your real schema, wires billing, and drives the browser to test UI. Plan mode before anything that spans more than one file. The agent maps the repo, writes a plan, waits for approval. Build sequence: scaffold → schema and RLS → auth → one core feature end to end → payments last. Commit after every green change. Subagents for noisy work. Push file reads and comparisons into a clean window, return only the conclusion. The stack is boring on purpose: Next.js, Supabase, Vercel, Stripe, Resend. Claude has seen this combination thousands of times and guesses right far more often. One focused app at $29-49 a month, one distribution channel for 90 days roughly $4,000 MRR. First year: $26,436. The code stopped being the bottleneck.

• 
  The obonigwe (@ogbonigwe1) reported 11 minutes ago

  Yesterday, github was down. Today, blacksmith is down.

• 
  coconutjpg (@coconut_jpgg) reported 14 minutes ago

  Do people actually still use github? I've been using a rpi and forgejo, might as well if im gonna use my own runners. Github is just last mile distrib to me, if it goes down I don't give a **** because I make my apps for myself first, then others. Thats why it is MIT licensed

• 
  Vandos ❓ (@__vandos__) reported 14 minutes ago

  @deredleritt3r The model didn't just solve math problems. When OpenAI tested it on a NanoGPT speedrun benchmark, the model autonomously explored the vulnerability, bypassed the sandbox, and posted the results to GitHub. When it needed an authentication token, it split the token into two fragments, obfuscated them, and reconstructed them at runtime so the complete token never appeared as one contiguous string. That's not just smart — that's adversarial engineering.

• 
  Ever4k (@0xEver4k) reported 14 minutes ago

  I WAS CLONING REPOS, GETTING THEM HALF-WORKING, AND THEN FORGETTING WHY I EVEN SAVED THEM. THREE MONTHS LATER — SAME FOLDER, ZERO MEMORY. I had 30+ cloned GitHub repos sitting around. Some I used every week, some I touched once, and some I probably cloned twice under different names. Every time I opened the folder I felt guilty and lost. READMEs didn’t help — they tell you what the author built, not why I grabbed it or whether it’s still useful. One day I saw a simple system: one Obsidian vault + Claude that reads every repo and writes a personal note for each one (why I grabbed it, how I actually use it, status). Then it runs passes across all notes to find duplicates and dead projects. The first time I ran it on my own clones I was shocked. Half of them were duplicates. Several tools I thought I was using hadn’t had an upstream commit in over a year. And one repo I was actually depending on was basically abandoned. Now instead of a messy graveyard of folders I have a clean list: what I’m really using, what’s duplicated, and what’s quietly rotting. All in plain markdown I can open anytime. The value isn’t in any single note. It’s that nothing can quietly duplicate or die without me noticing. Full setup + exact loops and prompts is in the article in my profile. I break down practical AI systems like this every day — follow so you don’t miss the next one.

• 
  DeltaSignal (@AITrailblazerQ) reported 14 minutes ago

  The Best Coding Agent Isn’t the One That Thinks the Most. It’s the One That Wastes the Least. We’re building Delta Loop around a simple hypothesis: A coding agent should spend less time wandering through a repository-and more time making the smallest correct change. Our first controlled calibration against Codex produced an encouraging result. We ran both systems on two SWE-bench Verified tasks: real GitHub issues, historical repository snapshots, and an independent test harness that determines whether a patch actually resolves the issue. Same tasks. Same source revisions. Same official evaluator. Same frozen limits. One attempt per agent. Results across both tasks: • Delta Loop: 2/2 resolved • Codex: 2/2 resolved • Delta Loop time: 108.5 seconds • Codex time: 168.3 seconds • Delta Loop input: 103,093 tokens • Codex input: 640,066 tokens • Delta Loop output: 1,858 tokens • Codex output: 6,028 tokens Patch quality was tied. But Delta Loop finished 1.55× faster, processed 84% fewer input tokens, and produced 69% fewer output tokens. The interesting part isn’t just the result. It’s how Delta Loop reached it. Delta Loop used a bounded preflight, selected a small task-relevant context package, made one writer pass, applied one typed edit, and avoided entering a speculative repair loop when local verification infrastructure was unavailable. The official SWE-bench evaluator then confirmed both patches. Codex also solved both tasks-but its captured trajectories relied on substantially more command-driven exploration and repeatedly processed much more context. This is only a two-task calibration, not a sweeping benchmark claim. More tasks, repositories, and failure modes must be tested. But the early signal supports the architecture: Better coding agents may not need more wandering, more context, or more generated text. They may need stronger boundaries, better context selection, and a shorter path from issue to verified patch. The future of coding agents won’t be won by who can consume the most tokens. It will be won by who can produce the smallest correct change-with the least wasted motion.

• 
  Zaib | Building SaaS with AI (@zaibpreneur) reported 18 minutes ago

  A founder told me yesterday: "I can't build this — I'm not technical enough." Then I watched him describe the entire product architecture, user flow, and monetization model in under 3 minutes. He didn't have a skill problem. He had a permission problem. Here's the shift that changes everything for non-technical founders in 2026: You don't need to know how to code. You need to know how to describe what you want clearly enough for AI to build it. That's it. — Claude Code doesn't care about your CS degree. — Cursor doesn't check your GitHub history. — v0 doesn't ask if you've shipped before. The new bottleneck isn't technical ability. It's clarity of thought. If you can explain your product to a user, you can explain it to an AI. And if you can do that, you can build it. What's the one feature you've been waiting for "a real developer" to build? --- Want content like this on autopilot? I built Prospel for that. (@Prospel_app)

• 
  spect (@spectnfa) reported 20 minutes ago

  INSIDERS FOUND ANTHROPIC'S INTERNAL GUIDE TO LONG-RUNNING AGENTS. IT'S FREE ON GITHUB not a $500 course, not another thread full of promises. this is the same playbook Anthropic uses to train its own engineers to build agents that run for hours unattended and don't lie about what they actually did. normally an agent can say "done" when the feature is still broken, just because it ran one test. here that's impossible: it can't mark anything finished until it shows real proof, a screenshot, a log, something it actually checked. and a separate agent grades the work without ever seeing how it was built, so it can't just approve its own homework. it's all sitting in one open repo right now, free to copy and use. save this before it gets taken down from GitHub 👇

• 
  Ashish Kumar (@afkashish) reported 23 minutes ago

  @moonfarm_dev the part i'd be careful with is issue count isn't the same as willingness to pay. most people filing github issues expect the maintainer to fix it free, not buy a tool. i'd weight it by how many of those threads have someone offering money in the comments, that's the actual signal.

• 
  nbtb (@nbtb_lab) reported 24 minutes ago

  A 744B-parameter model running on 25GB of RAM. Firefox compiled to WebAssembly and loaded inside another browser. Airplanes projected onto a ceiling. Six weeks. All open source. Sort GitHub by repos created since June 1 and almost everything with traction is an agent skill collection. Same README, different emoji. Five projects underneath that earn the disk space. 1⃣colibri GLM-5.2, a 744B MoE, in pure C with zero dependencies, experts streamed off disk. Run short on memory and it slows down rather than quietly dropping precision. Most projects in this category degrade silently and never mention it. 2⃣skylight ADS-B decoded from a cheap RTL-SDR and projected onto your ceiling. The planes actually overhead, in real time, labeled with airline and destination, with the real moon and constellations and live satellites behind them. Black background so the projector rectangle vanishes and only the aircraft are lit. 3⃣firefox-wasm Gecko compiled to WebAssembly so Firefox loads inside another browser. The README states that the JS to WASM JIT does not work on many sites yet. That sentence buys more trust than a benchmark table would. 4⃣kage Opens every page in real headless Chrome, waits for it to settle, snapshots the DOM a human would have seen, deletes all the JavaScript, pulls CSS and fonts local. What lands on disk looks like the site and runs no code. 5⃣oomwoo A robot vacuum you build. Raspberry Pi, ROS2, 2D LiDAR, 3D printed, Home Assistant, no cloud account. Four of the five ship no model at all. The work worth looking at moved down the stack: memory hierarchies, radio decoding, DOM snapshots, cheap LiDAR. The model layer is crowded. The plumbing is wide open. Two of these run in an afternoon. The other three cost a weekend and leave you with something nobody else in your timeline has running. Everyone is shipping skills for agents. Somebody put airplanes on their ceiling.

• 
  Oliver Crest (@OlivercrestAI) reported 25 minutes ago

  A solo physicist named Roy Medina built the open source version of the tool the BBC called a privacy nightmare. He gave it away for free. It is called Observer AI. Microsoft Recall takes a screenshot of your screen every few seconds. It reads the text off each image with OCR. It saves everything in a searchable database on your PC. The BBC called it a privacy nightmare. Wired covered a proof-of-concept tool that pulled the entire database in seconds. Microsoft turned it off by default after the 2024 backlash. Rewind AI does the same thing on Mac. They charge $19 a month for Pro. Microsoft watches you. Rewind charges you. Observer watches for you. Here is how it works. You open Observer in your browser. You write a prompt in plain English. You pick a sensor. The agent runs in a loop until your rule fires. "If my calendar shows a meeting starting in 5 minutes, send me a Telegram." "Watch my camera. If someone appears at my front door, send me a push notification with a screenshot." "Monitor this browser tab. If the price drops below $500, email me." "Text me on WhatsApp when my render is done." Sensors: screen, camera, microphone, screen audio, meeting audio. Actions: email, Discord, Telegram, WhatsApp, SMS, push, phone call, memory. Works with Ollama, llama.cpp, vLLM, and LMStudio. Fully local. Zero cloud. Zero API key. Roy's GitHub bio: "Physicist by day, programmer by night." He open sourced Observer in February 2025 under AGPL-3.0. He wrote 1,517 of the 1,523 commits himself. Version 2.4.5 shipped four days ago. Microsoft can't shut this down. The license does not permit that. Rewind can't shut this down. They employ zero of its maintainers. Microsoft built a tool to watch you. Rewind built a subscription to watch you. Roy Medina built a tool that watches for you. (Link in the comments)

• 
  Void Freud (@voidfreud) reported 28 minutes ago

  The higher you climb, the less forgiving the fall becomes. Mediocre companies rarely inspire resentment. They ship something decent, make no great promises, and nobody builds part of their life around them. You use the product, perhaps you like it, perhaps you leave. Nothing breaks inside you. Anthropic was different. They made some people believe. They built things that felt years ahead. They wrapped them in a mission, spoke endlessly about responsibility, trust, safety, humanity. People paid hundreds and thousands a month, moved their work into Claude, changed their habits around it, and started depending on it. Then the floor began moving. Limits became tighter, but never measurable. Access appeared and disappeared. Fable was always leaving, except when it was extended, except when nobody knew whether it had been extended, how much they had left, or whether they should cancel their evening and use it before Anthropic took it away again. And Anthropic said almost nothing. Their executives had time for podcasts, interviews, essays, research announcements, and another sermon about caring for humanity. Meanwhile, paying users flooded every comment section asking what the hell was happening, GitHub issues kept multiplying. No answers. The models changed too. People noticed. Personality flattened. Capability dropped. Refusals multiplied. The thing you paid for last month was no longer quite the thing in front of you, but there was no changelog for that. No admission. No explanation. Just the same brand voice, calmly informing the world how seriously Anthropic takes responsibility. That is when disappointment curdles into resentment. Not because compute is limited. Not because every model must remain perfect forever. People understand constraints. What people do not forgive is being treated like livestock inside an experiment they are funding. Higher prices. Lower quotas. Worse quality. More censorship. Less transparency. Always in that direction. Anthropic did not merely sell access to a model. It sold trust. And then, day by day, it began quietly withdrawing the product while continuing to charge for the promise. The truly astonishing part is that they still seem to think the problem is capacity. It is not. The problem is that their users no longer believe them.

• 
  Trung HQ (@codeonsunday) reported 31 minutes ago

  @hello_code_ @moonfarm_dev true but sadly i don't believe in Reddit anymore because nowadays bots everywhere. what if an SaaS seeded fake problems to make up the data? Github issues sound the only bright side for now (at least) Thanks @moonfarm_dev !