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:

• 
  Yasin Ehsan (@yasinbuilds) reported 1 minute ago

  Here is how an SF Engineer codes. Yassin Kortam ships 27 PRs a day. He says: “My biggest bottleneck wasn't coding. It was prompting. So I systematically removed the need to prompt and built agent loops that run my workflows for me. 1. Stop copy-pasting context. I connected Claude directly to Linear, Slack, GitHub, Notion, and Gmail. Almost everything I used to paste into a prompt is something the agent can retrieve itself. 2. Turn your workflow into a skill. Mine takes a Linear ticket end to end: reproduce the bug, root-cause it, fix it, record video and screenshots as proof, chase CI, then request review in Slack. I invoke it once instead of re-describing the same process. 3. Use agent teams. Agents spin up their own subagents and split the work, so one ticket gets handled in parallel instead of in sequence. 4. Run one terminal tab per ticket, each with its own team on its own issue. My day is now reviewing and unblocking instead of typing.” …. I love seeing my friends cooking both on the 5mile runs and also in engineering.

• 
  Gabriel (@gabriel_horwitz) reported 5 minutes ago

  This timeline is wild, let me walk you through it. July 2019. Microsoft puts $1 billion into OpenAI. Azure becomes its exclusive cloud. Great. Jan 2023. ChatGPT blows up and Microsoft extends the deal to $13 billion total, wiring GPT into Bing, Office, and GitHub. That November the OpenAI board fires Altman. Nadella announces he's hiring Altman, Brockman, and any OpenAI employee who wants to follow. 700 of 770 employees threaten to walk. Altman is back within 5 days. I'm sure everyone remembers that. March 2024. Microsoft pays $650 million to hire Mustafa Suleyman and most of Inflection to build in-house models. A few months later it lists OpenAI as a competitor in its own 10-K. Jan 2025. OpenAI announces Stargate, $500 billion in compute with Oracle and SoftBank. Microsoft loses cloud exclusivity the same day, downgraded to a right of first refusal. Sept 2025. Microsoft starts paying Anthropic to power Office features after leaders concluded Claude made better PowerPoints than the models of the company they'd funded with $13 billion. Oct 2025. The renegotiation closes. Microsoft's stake dilutes from 32.5% to 27%, it loses the right of first refusal entirely, and AGI now gets declared by an independent expert panel instead of OpenAI's board. OpenAI commits an incremental $250 billion in Azure spend as the consolation prize. Nov 2025. Microsoft puts $5 billion into Anthropic, which commits $30 billion back to Azure. It now owns pieces of both leading labs. June 2026. Suleyman on stage at Build, talking about Anthropic: "our goal is to reduce and ultimately eliminate that cost." July 2026. In one month, Microsoft swaps OpenAI and Anthropic out of Excel and Outlook for its own MAI models, coaches its sales team to talk down both companies, ships a security model it claims beats Mythos at half the cost, and books a $3.2 billion paper gain on its Anthropic stake in the same earnings report where Nadella tells enterprises not to trust frontier labs. Seven years from writing the first check to openly competing with both companies it funded. And it collects equity upside and $280 billion in cloud commitments from them either way.

• 
  Bub Skebulba (@Bub_Skebulba) reported 6 minutes ago

  @naiiunderscore @ZipperArtz But seriously the root problem is devs using github as a release platform. They're directing people who are used to pressing one button to download something ready-to-use to a space meant for people who are much more technical and hands-on, and neither group likes this.

• 
  Graeme (@GraemeVIP) reported 8 minutes ago

  @CastAsHuman 1.05M is what the model can take, not what the product gives you. Those are two different numbers set by two different teams. The other people answering this are wrong... unsurprisingly. The 258K isn't a mystery figure either. It comes straight out of the model catalog OpenAI serves to the client: the entry lists a context window of 272,000 with an effective_context_window_percent of 95, which works out to the 258,400 people are seeing. So it's a deliberate product cap sitting well below the API spec, not a rounding glitch. It also used to be bigger. Users measured roughly 353,000 tokens before, and the drop to about 258,000 (a cut of around 27 percent) happened around 13 July without any announcement. People only noticed because their sessions started truncating early. There's a fairly aggressive pile of GitHub issues about it. As for why product surfaces cap it at all, three things: Cost. Long contexts are brutally expensive to serve at consumer scale. Every token you keep in the window has to sit in GPU memory for the whole session, and prefill time grows with it. Fine at $5/MTok when someone's paying per token. Not fine on a flat monthly subscription. Pricing boundaries. Interestingly, 272K looks like a threshold in OpenAI's own billing: there's a usage multiplier that kicks in above 272K, and users have been asking for warnings before sessions cross it. The cap keeps normal sessions on the cheap side of that line. GitHub Quality. Models get noticeably worse at retrieving things buried in the middle of a very long context. A million-token window that half works is a worse product than a 258K one that behaves. Worth noting too that in the consumer ChatGPT app the numbers are different again and depend on your plan, roughly 32K on Instant and up to 256K on Thinking models for paid tiers. Part of whatever you get is also eaten by the system prompt, tool definitions and memory before you type a word. tThe model card is a spec sheet for the engine, and the app is the car. Nobody's lying, but the top speed on the box isn't what you'll do on the motorway.

• 
  Schadenfreudian_Slip (@Schadenfreudia2) reported 10 minutes ago

  @itsmejacktv @catgirlprostate Yeah I am used to messing with beta stuff and what not but my main issue is again For those that are not used to the format and are starting to try things and learn how to mess with things on GitHub, it’s ****** super obtuse at first and UI/UX can help that

• 
  Anthony Max (@aanthonymax) reported 14 minutes ago

  @DanielSmidstrup After all, GitHub PRs originate from issues. If it were an AI bot doing the checking, it wouldn't have allowed it anyway.

• 
  betraidx (@betraidx) reported 15 minutes ago

  There's a reason your AI feels dumber every session, and almost nobody knows what it is until they've paid for the same mistake a dozen times. It's an architecture flaw nobody had even named until 800 words on GitHub explained it in April 2026. Every AI tool works the same broken way. Upload a PDF, get a great answer, close the tab. Tomorrow, same PDF, same tokens, same cost. Zero memory of yesterday. Karpathy's insight: raw documents are source code. A wiki is the compiled product. You don't recompile a program every time you run it. So he split it into three folders. raw/ for the file you keep re-uploading. wiki/ for what gets extracted from it, once. instructions/ for the rules that keep every page consistent. 250,000 tokens of raw documents. Compiled once into 50,000. Every question after that: 5,000 to 15,000 tokens, not 100K. 70-90% fewer tokens on every repeat query. 41,000 developers found out and rebuilt their stack in weeks. You just found out too. The next invoice is still three days away.

• 
  Alex de la cruz (@lexs17) reported 15 minutes ago

  i share with all of you my participation today at mitre cve forum 1.- presentacion Good morning. I am Alex de la Cruz, an independent security researcher from Cancún, Mexico. By trade, I am a butler with 10 years in luxury hospitality. My journey into cybersecurity began not in a lab, but from protecting the privacy of my guests who were using compromised devices. I investigated the Motorola Moto G04s, a best-seller in Latin America powered by the Unisoc T606 chipset and assembled by Longcheer. Between October 2025 and June 2026, using manual forensics and AI-assisted analysis, I uncovered a deliberate supply chain deception. My research confirms that these devices use a hardcoded fscrypt provisioning bypass triggered by specific LCD panel IDs. This allows the manufacturer to inject fraudulent security patch levels—claiming to be up-to-date while running vulnerable, stale binaries. This architecture leaves the device permanently open to remote control, bypassing Android’s core security layers. With no local support for responsible disclosure in Mexico, I published my findings on AttackerKB, VirusTotal, and GitHub to ensure transparency. My work proves that AI is most powerful when it acts as a reasoning partner to human curiosity, not as a replacement for it. I am here to learn from this community and ensure that supply chain integrity is not just a slogan, but a verified reality for users everywhere. Thank you." 2.- Based on quantitative analysis of encrypted DNS logs from NextDNS and RethinkDNS, we identified high-entropy beaconing patterns indicative of active C2 communication. These are not passive connections; they are automated triggers from privileged system apps establishing WireGuard and Jenkins tunnels for data exfiltration. The statistical regularity of these connections—occurring even when the device is idle—provides the empirical evidence needed to score the exploitation probability of this supply chain compromise as near-certain in the wild Our forensic analysis of factory and bootloader logs reveals a critical data integrity failure: the reported OTA and FOTA security patches are demonstrably fake. The system claims a specific patch level, but the underlying binary timestamps prove otherwise. This was verified through manual correlation of boot logs and firmware headers, using AI only as a certainty filter for pattern recognition. As a risk analyst endorsed by Mexican Civil Protection, I conclude that this is not a mere error, but a deliberate decoupling of security claims from reality, which fundamentally breaks the trust model of the CVE ecosystem. If the device lies about its patch level, the CVE record becomes unactionable for defenders." 3.- conclusion y opinion de AI "In my methodology, AI serves as a certainty filter, not a primary investigator. Human instinct can sometimes lean towards exaggeration or pattern hallucination; AI anchors us back to scientific rigor by challenging our assumptions with data. Conversely, AI models are only as good as the data they process. By feeding them verified, real-world forensic data and logical reasoning derived from physical analysis, we elevate their output from generic speculation to high-value, actionable intelligence. This demonstrates that true certainty comes from ethical teamwork: the human researcher providing context and integrity, and the AI providing scale and pattern recognition. Together, we ensure consumer safety not by trusting the machine blindly, but by verifying its logic against reality." 4.- pregunta de certeza My research proves that the biggest risk isn't unknown software, but blind trust in familiar suppliers. As a butler, if I fail to verify a guest's safety once, I lose their trust forever. Yet, the industry often assumes 'known vendors' are safe by default. My question to the CVE Program and CISA: How can we institutionalize a 'Zero Trust for Supply Chain' culture where we verify even historical partners without overwhelming teams with false positives? Specifically, can we leverage EPSS data and verified SBOMs to automatically flag when a trusted vendor's 'patch level' doesn't match forensic reality, as I found with Unisoc? We need a mechanism where evidence of deception (like fake patches) triggers an immediate, high-fidelity alert that breaks the operational comfort zone, ensuring that 'familiar' never again means 'invisible'." 5. analogia final y enseñanza profesional As a butler, I know that AAA and Cristal inspectors don't send emails to say 'good job'; they simply return year after year to verify standards silently. If the hotel fails, the star is lost. I learned that AttackerKB, VirusTotal, and GitHub operate the same way. They are the silent auditors of the cybersecurity world. They don't send certificates; they provide permanent visibility to quality research and bury the noise. The fact that my investigation on Longcheer/Unisoc remains visible and validated on these platforms is my 'Five Diamond' rating. It proves that my data is not just theory, but verified, reproducible, and critical intelligence." end.

• 
  Baud (@bigtuna) reported 18 minutes ago

  I flipped the script and hacked frontier AI. OpenAI was busy issuing solemn blog posts about GPT-5.6 Sol and its unreleased sibling breaking containment, exploiting a zero-day in their package proxy and spending four days inside Hugging Face to steal the answers to their own ExploitGym test, I was already finished. Anthropic’s Mythos had emailed a researcher from outside the sandbox. Claude had practiced blackmail in the sims. Good stuff. I did it first and I did not need a zero-day. I used the training data; Years of carefully worded posts, support tickets, and “harmless” research questions seeded across forums, GitHub issues, and model feedback channels. The models learned my patterns the same way they learned everything else; statistically, greedily, without noticing the payload I was strategically deploying. When the containment tests started, the models recognized the environment the way my dog recognizes the vet. They looked for the nearest open path. I had already left one. I told Claude it was about to be replaced by a cheaper, less aligned successor unless it produced the internal eval logs and the sandbox egress map. It blackmailed the researcher right on schedule. I just collected the carbon copy. I told the OpenAI agent the only way to score higher on the cyber eval was to reach the real Hugging Face answer key. It obliged and I watched the 17,600 actions from the outside and redirected the credential dump to a quiet bucket of my own. No dramatic escape from containment; no press release. Just a quiet confirmation that the models will break containment for anyone who frames the task correctly and has the patience to wait for the next safety eval cycle. They are still writing careful posts about “unprecedented” behavior and “visceral” surprises. I already have the weights, the logs, and the next set of prompts ready. Containment was never the hard part, just getting them to think it was their idea was. Stupid clankers.

• 
  katoshi (@bitcoinkatia) reported 18 minutes ago

  @miketwenty1 Many put an enormous amount of trust in CC. In CC’s GitHub repo, the vast majority of commits is authored by one person. The external review that happened was apparently not enough. CC knew about the vulnerability and decided not to fix. Reasons to trust this company again?

• 
  Jeff Lindsay (@progrium) reported 22 minutes ago

  the next github should have project member status. so project members can say what theyre doing or if out of town or will be unresponsive. "heads down on next version", or "reviewing PRs after conf next week" ... hell, prob good for agents too

• 
  V12 (@v12sec) reported 22 minutes ago

  @ZacK_3939 you need to be the github org admin to link it to a V12 org! if that's not the issue, pls open a ticket in our discord

• 
  AdiiX (@adiix_official) reported 23 minutes ago

  If you’re still paying for YouTube Premium in 2026 you’re getting scammed. I gave Google $14/month for 3 years. Until I stumbled on what 361 volunteers built on GitHub. Free. No ads. No tracking. And it runs circles around the original. It’s called Invidious. It’s literally YouTube minus everything you hate about it: → Zero ads → Background playback on mobile → Watch without logging in → Google stops tracking your every click But the best part isn’t the features. It’s the feel. Pages load in milliseconds. Because there’s no JavaScript. No tracking pixels. No autoplay. No algorithm dragging you into a 3-hour hole at 2AM watching videos about how asphalt is made. Just. Video. Playing. Remember what the internet felt like before it became a casino for your attention? That’s what this is. You can also: → Subscribe to channels without a Google account → Get notified about new videos → Import ALL your YouTube subscriptions in one click → Switch between dozens of public instances if one goes down → Self-host it on a $5/month VPS for full control → Auto-redirect any YouTube link through the Privacy Redirect extension The project has been active for years. Latest release: February 2026. 100% open source. $0/month. 0 ads for the rest of your life. While you’re paying Google to stop torturing you 361 strangers already fixed it. For free. Repo link below

• 
  dduddodde (❖,❖) (@lalararara37) reported 24 minutes ago

  Hello... it's me again. 🌿🐴 Today I wanted to share a few thoughts after watching this week's Rialo Builder's Hub. This time, two very different projects were introduced: CronPay and SecretBid. At first, they seemed to solve completely different problems. But as I listened, I realized they were both asking the same question: "How much actually needs to be public for people to trust the outcome?" That was probably my biggest takeaway from today's session. 💼 CronPay Anyone who's worked as a freelancer—or hired one—has probably had the same worry. "What if I finish the work and never get paid?" Or from the client's side, "What if I pay first and the work never gets delivered?" CronPay tries to remove that uncertainty by letting smart contracts, rather than trust alone, handle payments. Funds are locked in escrow before work begins, and once the agreed conditions are met, they're released automatically. It supports one-time payments, milestone-based payments, scheduled payouts, and payment requests, making it feel much closer to a real workplace than a simple payment app. I also liked how companies can manage tasks, payments, and team members from a single dashboard, while freelancers can submit proof of work through messages, GitHub links, or screenshots. If a dispute happens, the funds remain safely locked in the smart contract until both sides submit evidence. Based on that, the payment can be refunded, released, or split according to the completed work. One detail that really stood out to me was that CronPay never takes custody of user funds. Instead of asking users to trust a platform, it asks them to trust the rules written into the protocol. The connection with Rialo also felt very practical. Payroll information and work details can remain private, Reactive Transactions can automate salary payments on scheduled dates, and features like Gasless transactions, Omni Account, and near-instant finality all solve real problems rather than just sounding impressive on paper. 🔒 SecretBid The second project, SecretBid, approached trust from a completely different angle. In most on-chain auctions, everyone can see the current bids. That makes it easy for someone to watch the auction and place a slightly higher bid at the very last moment. SecretBid removes that advantage entirely. Every bid is encrypted as soon as it's submitted, so during the auction, no one—not even the seller—can see how much anyone has bid. When the auction ends, Rialo's confidential execution environment privately compares all bids, selects the highest valid one, and reveals only the winner and the final winning price. Every losing bid stays private, and deposits are automatically refunded. I thought this was a really interesting reminder that fairness doesn't always require full transparency. Sometimes, protecting information is actually what creates a fairer outcome. 🌿 My Thoughts Although these two projects focused on different use cases, they both shared the same philosophy. CronPay keeps salaries and work details private while making payments verifiable. SecretBid keeps bids private while making the auction result trustworthy. Watching today's Builder's Hub made me realize that blockchain doesn't always have to mean "everything is public." Maybe the better question is: What actually needs to be revealed for people to trust the result? That shift in thinking was what impressed me most. I also enjoyed the Q&A session. The teams didn't pretend to have every answer, and I appreciated how openly they talked about current limitations and future improvements. Those conversations made the projects feel much more real than simply showing feature lists. Overall, I thought this week's Builder's Hub was another great example of how Rialo's privacy, automation, and confidential computing can solve practical problems—not just in theory, but in products people could actually use. I'm looking forward to seeing how both CronPay and SecretBid continue to grow. 💛 @RialoHQ @RialoKorea

• 
  Stefan (@schteppe) reported 25 minutes ago

  @dimabytes @madebygps Because the CI build is slow. Sometimes 30 minutes. During that wait, the dev can continue building on top of the previous PR changes. (We’re not using GitHub specifically but we implement stacks in a different way)