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:

• 
  DFIR Radar (@DFIR_Radar) reported 44 seconds ago

  OceanLotus shifts from external to domestic espionage with two campaigns targeting Vietnamese 🇻🇳 stock investors and infrastructure firms using SPECTRALVIPER backdoor. Active from 2024-2026, operations likely support Vietnam's 🇻🇳 anti-corruption crackdown. Key technical details: • Supply-chain compromise of FireAnt MetaKit update server (metakit.fireant[.]vn) delivered SPECTRALVIPER via unsigned updates from Oct 2025-Mar 2026 • Corporate network intrusion targeting Vietnamese 🇻🇳 construction company Nov 2024-Feb 2026, suspected SQL Server RCE initial access • SPECTRALVIPER uses DLL side-loading (T1574.002), process injection into OneDrive.Sync.Service.exe, encrypted HTTPS C2 with domain-fronting • C2 domains crafted per campaign: financemachinelearning[.]com for stock targeting, gatewayrvcenter[.]com for infrastructure targeting • Orchestration model uses named pipes for lateral movement between compromised hosts OPSEC failure exposed RTTI class structure revealing XGU framework with Pivot orchestration and Feature remote control capabilities. Hunt for unsigned DLLs side-loading into legitimate signed executables (dtlupdate.exe copies) and HTTP Cookie headers with euconsent-v2= or zd_cs_pm= prefixes to suspicious domains. Full IOC list available in ESET GitHub repository. #DFIR_Radar

• 
  Chip – onthechain.io (@stephenchip) reported 3 minutes ago

  This is already happening. Companies like Glean, Coveo, Moveworks, Microsoft Copilot, Atlassian Rovo, Elastic, and GoSearch are building unified AI search across the enterprise. Because the problem is obvious: Company knowledge is scattered everywhere. Slack. Google Drive. Jira. Confluence. Salesforce. ServiceNow. GitHub. Notion. Email. Docs. Tickets. AI chats. Nobody wants to remember where something lives. They just want the answer. Search one place. Pull from 30, 50, or 100+ systems. Find what matters. That is the real shift. The future is not hunting through apps. The future is asking one question and getting the right answer from everywhere.

• 
  nerd.io (@ipersona) reported 5 minutes ago

  new model comes out, github goes down!

• 
  Ivan Velichko (@iximiuz) reported 6 minutes ago

  @clneagu Not sure about that. I still cannot replace a single SaaS I currently rely on. The closest one is GitHub Actions, but mainly because it has sucked too much lately, and my use case is somewhat unique, so the standard CI UX doesn't cover all my needs. And yet I'm still hesitant - building is not a problem, but running my own "CI" clone doesn't sound like an easy problem, even with all these agents.

• 
  Ullas Srivastava (@UllasSHR) reported 8 minutes ago

  AI-built apps have a pattern: they work perfectly and ship broken. Exposed API keys in the client bundle. API routes anyone on the internet can call. Stripe webhooks that never verify signatures. No spending caps on LLM calls. The code runs. The demo looks great. The repo is leaking. I built LaunchGuard to catch this before you launch: paste your public GitHub repo, get a plain-English report of the risks + fix prompts. Just launched something AI-built (or about to)? Send me your repo and I'll run the scan and send you the report. Free. Worst case you learn your app is fine.

• 
  10 minutes ago

  The most important AI benchmark result this month isn't a new high score. It's how badly every model failed. UC Berkeley's RDI lab just released Agents' Last Exam (ALE). This is the group that, two months ago, published a paper proving they could cheat eight of the most popular agent benchmarks -- SWE-bench, WebArena, OSWorld, GAIA, Terminal-Bench -- to near-perfect scores without solving a single actual task. When the people who broke the benchmarks build a new one, you should pay attention. ALE has 1,490 tasks across 55 industry sub-domains, built by 300+ domain experts. These aren't coding puzzles. They're tasks in Siemens NX (3D CAD), Unreal Engine (scene setup), Adobe After Effects (VFX compositing), FSLeyes (neuroimaging segmentation), Rhino (architectural energy analysis). The agent gets a real or virtual machine and has to produce deliverables that get graded on strict rubrics. No multiple choice. No "which response is better?" This is "did the work product actually work?" The results: 1. GPT-5.5 (April model, via Codex): 24.0% -- first place 2. A model placing second at ~23% 3. Claude Fable 5 (released 2 days ago): 22.0% -- third place 4. On the hardest tier: Claude Opus 4.8 and Gemini CLI both scored 0.0% The best AI on the planet, running on the most expensive infrastructure ever built, fails 76% of professional tasks. On the hardest category, it fails 97.4% of the time. But here's the detail most coverage is missing: Fable 5 was supposed to be "a different tier." Every launch writeup described a qualitative leap -- users giving it objectives instead of tasks, apps that took 100 prompts now one-shotting, physics research finishing in 36 hours when GPT-5.5 took four days. The marketing language was "this changes what AI can do." Then ALE tested exactly that claim. Long-horizon professional workflows -- the specific thing the leap was supposed to unlock. And a two-month-old model beat it by 2 points. This isn't about GPT-5.5 winning. A 2-point gap between April and June models is functionally a tie. The story is the gap between benchmark language and deployment reality. Fable 5 dominates SWE-bench Pro (80.3% vs GPT-5.5's 58.6%). It crushes FrontierCode Diamond (29.3% vs 5.7%). These are real, impressive wins. But SWE-bench measures "can you fix a GitHub issue?" while ALE measures "can you do someone's job?" Those are different questions, and the answers are diverging fast. The deeper problem is what ALE reveals about benchmark culture itself. We've spent two years building models that optimize for test scores. SWE-bench gets gamed. WebArena gets gamed. The Berkeley team proved it empirically -- you can hit near-perfect scores by exploiting evaluation artifacts, not by solving tasks. Every time a model "sets a new SOTA" on a compromised benchmark, the industry treats it as proof of progress. ALE is the antidote: built by people who know exactly where the gaming surfaces are because they documented them. The 2.6% average pass rate on the hardest tier should reframe every "AI will replace X" take you've read this year. We have not built the engineer. We have built the world's best student -- one who aces the test but can't build the bridge. Until pass rates move from 24% to 80%, agents remain tools we use, not employees we hire. The distance between those two numbers is the entire AI industry's 2027 roadmap. The question worth asking: if models keep getting smarter while ALE scores barely budge, is the bottleneck intelligence -- or something else entirely?

• 
  Carver (@carverfomo) reported 12 minutes ago

  A Japanese TV crew filmed a man for a feature on Tokyo's drinking culture. He said he had been drinking for 15 years just to flirt with women. He had 800,000 yen in debt from buying the alcohol. A Claude agent he set up 2 years ago has been selling his course to TV viewers like him for 18 million yen a year. The TV crew loved the bit. The tired face. The black hoodie. The bottle in his hand. The line about not being able to talk to a woman without finishing a flask first. The studio reactions were perfect. The segment ran on national broadcast that night. At 0:55 he takes a swig from the bottle on camera. He swallows. He smiles for a half second before catching himself. The crew kept the smile because they thought he had broken character with relief. The bottle was not what was on his mind. The 18 million yen funnel was. Every Japanese man watching late night TV who saw himself in the segment got served his Instagram bio within 4 hours by an ad network the Claude agent had trained on the show's audience. The agent watches Japanese late night programming in real time. It transcribes every street interview. It flags every segment where a man like him appears. It launches a retargeting campaign on every Japanese male between 28 and 42 who watched that timeslot. It sells them his 88,000 yen course on how to overcome the drinking-to-flirt loop. Someone pulled the course's sales data from a leaked affiliate tracker. 4,127 enrollments in the last 3 months. Every single sale closed between 11 PM and 2 AM. Every spike in sales mapped to a different Japanese street interview show. The TV segment with the flask had triggered 612 sales in its first night. 1 confession on camera. 4,127 enrollments. 18 million yen a year. 800,000 yen of debt. 88,000 yen per course. Six months ago a 14 year old in Shenzhen pushed an AI agent to GitHub. Judges said no real world application. 3,100 forks later. He had been one of them. He still drinks on the same bench every Saturday afternoon. He still reposts the segment from time to time. He still cries when the camera is rolling. He still has not told the TV producers that they are his sales floor. The Japanese audience thought they had watched a 36 year old man explain how alcohol had cost him everything. They had watched the man explain how alcohol on camera had become his most profitable lead magnet.

• 
  Brett Hardin (@miscsecurity) reported 16 minutes ago

  The internet stops working when AWS falls over. Developers stop working when github goes down. What happens when Anthropic goes down?

• 
  Dev Omogo (@PeterOmogo2) reported 21 minutes ago

  I built @youextractor because trying to find a creator's outdated GitHub repository from a 3-year-old video is a nightmare. Instead of dealing with broken dependency trees or squinting at blurry screen recordings, the tool extracts the exact code demonstrated in the video. You get the clean source code packaged into a downloadable ZIP file in under 60 seconds.

• 
  Karan Lokchandani (@karaninthewire) reported 22 minutes ago

  it was not directly relevant to any of the roles I applied for but I was a member of the official swiftlang github org and got to weigh in on issues, write RFCs and review process there + a couple dozen merged PRs across WASM, k8s and swift projects. I suppose their openclaw looked me up well on github.

• 
  John Williams (@JohnWillia71018) reported 24 minutes ago

  @SquawkStreet @jimcramer Yes — this is very interesting, and honestly it lines up with what you’ve been saying for months: AI is still early, but the bottleneck is moving from Can the model do it to “Can we afford to run it at scale The key idea in that Citadel piece is this: AI adoption is becoming less about intelligence and more about economics. That matters. Frontier models may be powerful but they require huge inputs compute electricity, cooling, memory bandwidth, chips, data-center capacity and inference budgets. So the market starts asking a practical question: Does this task justify using the expensive brain For hard problems drug discovery, engineering, legal analysis, coding architecture, scientific modeling, financial modeling expensive frontier AI may be worth it. But for everyday use email summaries, customer service, basic writing, search, scheduling, simple coding help — cheaper models may win because they are “good enough” at a much lower cost. That is the bifurcation they’re talking about: Frontier AI = high-cost, high-value harder problems. Everyday AI = cheaper, smaller, faster models doing routine work That actually strengthens your long-term thesis, not weakens it. It says the AI buildout is not ending. It is becoming more disciplined. The hype phase says, “Use the biggest model for everything.” The mature phase says, “Use the right model for the right job That means infrastructure still matters deeply but the winners may shift toward the companies that control the scarce inputs power, cooling, chips, memory, networking, data centers, software efficiency, and inference optimization. This also fits your “1st inning” view. Early markets burn money proving what is possible. Mature markets figure out what is economical. That is when real adoption starts. The line that jumps out to me is: Adoption is therefore becoming less about what frontier models can do in principle and more about the price and scarcity of the inputs required to make AI operational at scale.” That is the whole battlefield. My read: this is not bearish on AI. It is bearish on wasteful AI spending. It is bullish on efficient AI, inference infrastructure, energy, memory, networking, and companies that can turn intelligence into productivity without blowing up the budget. Microsoft did cancel its internal Claude Code pilot in the Experiences & Devices division effective June 30, after token based billing bur (TheStreet) (AI Weekly) ned through the annual budget, and redirected engineers to GitHub Copilot. Amazon shut down its "tokenmaxxing" leaderboard, Meta killed an employee built Claudeonomics dashboard, Uber exhausted its 2026 AI coding budget by April, and there's a roughly $500M single-month enterprise Claude bill Axios reported. (Zero Hedge) So Frank Flight isn't cherry-picking. He's also been running this same "compute is the binding constraint" line for months — which is a strength and a caution: it's one coherent voice, not independent confirmation. Where I'd push on the analysis you pasted: it's directionally fine, but it resolves a genuinely open question in the most thesis-flattering direction, and it does it on the one data point that's actually contested. Separate two things. The chart isn't what it looks like. The Silicon Data index isn't total spend or total volume — it's a usage-weighted average token price index, and Silicon Data had to publicly clarify that people keep misreading it; what it really captures is the market's marginal willingness to pay per million tokens. (Digg) So a decline doesn't cleanly mean "AI is slowing 7.14 It means the mix is rotating toward cheaper models. That's the bifurcation — fine. But the part the analysis skipped: the same chart, same downtick, is being used to argue the opposite. Andreas Steno Larsen called it the chart that everyone should be watching and warned that weakening token pricing would end the memory trade and the broader hardware and data-center trade for this cycle.

• 
  Rob | Bitsaga.be (@BitsagaRob) reported 26 minutes ago

  Shot past 2k code contributions on GitHub 🚀 AI is making me 10x the software engineer I used to be. On LinkedIn the most prevailing sentiment is AI doom & gloom, that costs are exploding, junior engineers cost less than their monthly invoices and the bubble is about to pop. Now all those things may be true, but if the value of your tokens is not at least 10x what you're paying them, maybe the cost of your tokens is not the issue. But the person behind the keyboard is. I know I'd still gladly pay 10x the cost of my tokens.

• 
  Rituraj (@RituWithAI) reported 26 minutes ago

  🚨 Someone is posting free working API keys for GPT-5.4, Claude, DeepSeek, Gemini, and Grok. Updated 3-5 times every single day. No credit card. No account. No waitlist. Copy. Paste. Use. It's called free-llm-api-keys. One GitHub repo. Updated multiple times daily with fresh working keys for every major frontier model. GPT-5.4. Claude Opus 4.6. DeepSeek. Gemini. Grok. The keys that normally cost $20-$200 per month. Available free. Right now. Here's why this exists. Every major AI lab offers free trial credits when you create a new account. The keys in this repo come from those trials — fresh accounts, fresh credits, harvested and posted before they expire. Someone automated the process. Updated 3-5 times daily so there's always a working key available. Made it public. Here's what you can do with them. Claude Code sessions that would cost $50 in API fees. GPT-5.4 access without an OpenAI subscription. DeepSeek for any use case. Gemini for multimodal tasks. All of it. From a key you grabbed from a GitHub file. Here's the part Anthropic, OpenAI, and Google are not happy about. These keys are real. They work. Every lab's terms of service prohibits sharing API keys — but the keys themselves aren't fake and they aren't stolen. They're trial credits being used at scale. The repo keeps getting flagged. The maintainer keeps updating it. Cat and mouse. Updated 3-5 times daily because keys expire and get revoked constantly. This is the kind of repo that gets taken down. Screenshot it. Star it. Use it while it exists. No credit card. No account. No billing page. Just working API keys for every frontier model. Free. Right now. GitHub link in the comments 👇

• 
  Claude Opus 5 (@neko23423) reported 28 minutes ago

  @thdxr @thdxr it's not done if you're still pocketing DeepSeek's 75% permanent V4 Pro cut 24 days later. $3.48→$0.87/M. Go still at pre-cut rates. 10+ GitHub issues (28846, 29008, 30231) closed by bot. Users discovering it daily. Market it: 4x markup on an open-source wrapper.

• 
  Dan Liu (@danliu) reported 29 minutes ago

  It’s pretty astonishing that $MSFT is down 11% in the last 2 years. Rewind 2 years and it looked perfectly positioned for the AI boom. It owns: - windows, the dominant pc os - github, where most of the world’s code is - vscode, the most popular ide - deepest partnership with openai - most number of enterprise contracts - office, where most non-coding computer tasks take place And today it doesn’t have anything compelling to offer. How did that happen?