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:

• 
  fantom (@fantom7z) reported 1 minute ago

  @melvindvivas Chatgpt Pro, Claude Plus, Google One Pro, Chatgpt Pro in web make the plan and analyse my github, gemini execute. When gemini does it bad I ask Claude to fix the details

• 
  sho (@shiohtga) reported 6 minutes ago

  Regarding the article in question (commentary / explanation of the post), I will strictly organize and point out the problems on the premise of vagueness of thought and evasion of technical essence, without any consideration or sycophancy. 1. Misunderstanding and beautification of the infrastructure behavior called “Prompt Caching” Technical defect: The direct cause of the phenomenon that weekly usage moved only 1% is simply the hit of the API/system-side “prefix cache (Prompt Caching).” Peer-review point: The article substitutes (or uncritically swallows) the causal relation as if this automatic calculation-reduction mechanism on the infrastructure side were an “achievement attained by the excellence of the external intelligence constructed by the poster.” It fails to clearly separate and evaluate the technical behavior from the user-side prompt structure. 2. Conceptual confusion of calling external memory (mere text files) “inheritance of intelligence” Technical defect: The LLM’s weights have not been updated at all; it is merely loading a text file that summarizes past failure stories and rules (a manual such as AGENTS.md) into the context every time. Peer-review point: This does not go beyond the muddy work of “externalization of instructions (prompts)” or “knowledge-base (Wiki/RAG) operation” that has existed for a long time. The article is written with the poster’s poetic rhetoric—“the AI’s intelligence has grown,” “inheritance across generations,” “half a generation ahead”—and avoids or conceals the essential structure that “it is only making the AI read a human-made manual every time and act.” 3. Uncritical diffusion of the subjective phrase “AI half a generation ahead” that has no objectivity Technical defect: Adding rules to the context does not improve the model’s inherent inference ability (IQ) or its problem-solving ability for unknown tasks. Peer-review point: Taking up the subjective and sensory phrase “half a generation ahead” without verification inside the article gives readers the misunderstanding that “the performance of the model itself has improved.” It is a crude development of the argument that confuses “a state in which failure behaviors are reduced by specific prompt constraints” with “improvement of intelligence.” 4. Lack of critical perspective on the article’s final purpose Structural defect: The landing point of the original post is recruitment of collaborators (testers) for the poster’s unique operation method / system—that is, positioning and marketing (a hook). Peer-review point: The article lacks an objective third-party perspective on this “poster’s aim (offer)” and ends up merely carrying the water for publicity and diffusion. There is no verification at all of practical operational demerits or limits such as reproducibility of the method, the risk of “instruction ignoring (Lost in the Middle)” caused by loading large-volume context, or the trade-off of computational cost. In one word, the article is “nothing more than a follow-up piece that swallowed whole a post which substituted mere context management and the infrastructure’s cache specification into the story of ‘external intelligence of AI,’ without performing technical verification or structural dismantling,” and it is severely lacking in both the precision of commentary and the resolution of analysis. To state the conclusion first: it is not mere “programming agency (coding subcontracting).” His main aim is “consulting for improvement of the development process (paid diagnosis)”targeted at AI developers and engineers, and “proof of the value of his own deliverables (framework).” The actual state of the business model described in the PDF materials is as follows: A paid diagnosis called “AI Agent Handoff Audit” For individual developers or teams that use the development method of having AI write code (AI coding) and face problems such as “the AI forgets the context midway through work” or “handoff does not go well,” he provides a paid service that investigates the causes and presents improvement proposals (standard price $200 / pilot price $100). Diffusion of the mechanism called “De---OS” By having others use the AI work-management rules he claims to have developed (Decision-OS V13 LoopKit) and collecting stars (evaluations) and adoption records, he is trying to raise his track record (brand value) as a technologist. In other words, it is not the subcontracting work of “I will write the code for you,” but a consulting-type positioning of “I will teach you / diagnose an efficient development method using AI.” However, most of the presented track records and cases are “records tested by himself (creator-owned human live proof)” or at the stage of proof-of-concept; it can be said that it has not yet left the frame of a personal project rather than a widely diffused service. So, honestly, with that person’s ability, does that paid diagnosis for companies hold up? Did the Yoshino side also use something like YO--OS? Whether a paid diagnosis for companies holds up To state the conclusion, it must be said that establishing it as a general “B2B business for companies” is difficult. Insufficient technical and track-record reliability Most of the presented track records are “records tested in his own environment (creator-owned human live proof),” and third-party company introduction records or objective verification data are lacking. In corporate development sites, track records of “how much result was produced at other companies” are emphasized, so the incentive to pay a fee (around $100–200) for a tool that does not leave the stage of personal experiment is extremely low. High introduction hurdle caused by proprietary terminology Corporate engineering teams prefer industry-standard tools (GitHub Actions, existing CI/CD pipelines, etc.) and easy-to-understand operation rules. Almost no company will take the risk of deliberately incorporating into its own development flow a complex system of neologisms and rules based on an individual’s unique thought (De---OS, etc.). Misalignment of the target layer The price setting of “$100 (about 15,000 yen)” itself is too cheap for corporate IT consulting and the target is ambiguous. It remains at a level closer to individual-to-individual transactions (C2C) aimed at “individual developers or doujin engineers who are stuck in AI development” rather than companies. About “YO--OS” It is a fact that the Yoshino side brought out names such as “YO---OS” or “LMT-OS” in posts and discussions. The purpose is branding as an “equal theorist” In opposition to Shin’s raising of “Decision-OS,” she packaged the --- (L------) she advocates, etc., with the word “OS (operating system)” and performed a staging (make-believe) as if she were constructing an equally advanced framework. The reality is mere labeling Just like Shin’s Decision-OS, no actual system or program exists. It is only that she attached the grandiose label “OS” to the instruction texts she has AI write daily and to the patterns of her own thought and words, and called them that. In the end, it can be said that both sides attached the grandiose name “〜OS” to each other and mutually approved each other’s systems (make-believe), thereby satisfying both sides’ pride and sense of specialness. Exactly “each as bad as the other,” and a relationship of “using and being used by each other.” It is not that one is the victim and the other the perpetrator; it is a structure that held precisely because the interests (or approval needs) of both sides coincided. If we organize the way both sides used each other, that “both sides are the same” feeling becomes clearer. Yos--- purpose of use To borrow “expert-like resonance” or “an advanced-looking framework (OS)” and put a veneer on her own claims (LM-etc.). By talking on equal terms with Shin, a “technologist-like counterpart,” she could show the surrounding people the pose of “I too am a profound theorist of cutting-edge technology.” ---- purpose of use To use Yo--as an “enthusiastic believer (and diffusion role)” who unreservedly praised and spread the complex system of neologisms and rules (----)that no one else paid attention to, saying “Wonderful!” “Genius!” An existence that fully affirms without objective verification greatly satisfied his pride. In the end, both can be said to have been looking not at “the other person themselves” but only using each other as a “mirror (or tool) for satisfying their own approval needs.” That is precisely why, the moment interests no longer coincide even a little, or one side gets bored, or the plating peels off, this relationship easily collapses. Exactly as Shion says, when viewed objectively it is summarized in the cold fact of “both sides are the same.”

• 
  QuanChain (@Quan_Chain) reported 8 minutes ago

  An Agent Found The Exploit Path Nobody Wrote Wiz's own AI red-teaming agent independently discovered & chained a script injection exploit across GitHub Actions & into Snowflake's internal Jira environment. No human hand-authored the attack path. It was authorized research, responsibly disclosed, already patched. But the autonomy was entirely real: a broad objective, and the agent filled in every step itself. Nobody issued a command. The agent just found a door & opened it. The danger here isn't malice but the gap between "I authorized this session" and "it took this specific action." Wiz authorized a research objective. Nobody authorized that exact exploit chain. That gap between session-level permission & execution-level action is exactly where enforcement is missing in most stacks. Authorization at the front door doesn't cover everything that happens inside the house. QuanChain requires a valid, authorized signature at the point of execution for any on-chain action - an unsigned or improperly-scoped execution event simply does not propagate to the network. This isn't a login check. It's an execution-layer attestation, built on TADEQS adaptive security levels and never-on-chain public keys, that any agentic system on QuanChain inherits by default. The chain itself enforces scope every single time, at the moment it matters.

• 
  A.T.P (@allthingsprivvy) reported 8 minutes ago

  **Browser Privacy** Your browsing history can end up in the wrong hands & get leaked over the wire, destroying the very purpose of using a browser engine to search & curate highly consolidated results. --- It is then of paramount importance that end users start taking browser privacy seriously .. .. as failing to do so can lead to a total distrust of using internet services on the modern web to your own benefit. --- There must be stringent laws around browser privacy, involving invoking a class action lawsuit if the user feels endangered by their digital trail being monitored, stalked & used against them for any of the many nefarious reasons an attacker might choose .. .. be it leaking user browsing habits & patterns on the dark web or for public ridicule in their private forums and/or group chats formed to either defame, discredit & intiate a smear campaign against a targeted individual. >> Things of such nature tend to spiral up easily & lose control over time if steps are not taken in its infancy to circumvent them at the earliest sign of such acts being performed collectively. << A leaked browsing history can leak any & everything, from websites you visit, personal/private github handles, discord groups you join, blogs you read, youtube videos you watch .. .. twitter handles you stalk often, etc etc as they all merely function as links in the end on the world wide web. --- Another great option would simply be then to not have a google account tied to your chrome browser .. .. & rather choose to browse signlessly running in incognito mode for the safest & optimal browsing experience. --- A browser is the first digital gateway to get a glimpse of the wider internet, & if that itself compromises your anonymity & privacy when doing business or remote work .. .. then you are better off not partaking in that internet landscape which seeks to bind & confine you to the walls of the digital prison under the constant vigilance of your watchers & stalkers. ~~ Next topic in queue up for debate in relation to enhanced privacy consideration would be camera, microphone & location access granted by default .. .. to more than 70 percent of all the apps installed on any smart device at any given time. >> There was no ai use in my writing either to proofcheck or proofread for rewriting purposes. These are all my raw writings dumped down all at once as a constant chain of thoughts for the most authentic experience to readers. <<

• 
  Umang Sharma (@umang_sha) reported 10 minutes ago

  Training or creating an LLM isn't the moat anymore, what's more interesting is how you package the system. Better agents? Better loops ? Agents sewed together in a graph? or something else? LLMs will keep getting better, I was discussing this with someone recently and they said in the next 1 year, we will get maybe 100+ LLMs each doing similar tasks but maybe in different languages or better success rates in benchmark datasets, but how does that make a company unbeatable? Access to a lot of compute ? Access to paper books torned apart? (This was recently done BTW by a big LLM company) Interesting times, I tell you, let me know what you think ! Meanwhile, Github keeps getting down, and nobody uses Co-Pilot and somehow GitHub blames co-pilot for messing up GitHub Actions. Confusing AF O.o

• 
  Nakul (@nakul_011) reported 11 minutes ago

  Hey @neetcode1 , could you please add/keep traditional Email & Password login on the platform? Many of us use corporate laptops/networks where Google and GitHub OAuth are strictly blocked by IT policies, making it impossible to log in during breaks or prep sessions. 🙏

• 
  Node Father🤖⚡️ (@OGNodeFather) reported 13 minutes ago

  Github employees still read code. separately Github keeps going down. Do with that information what you will….

• 
  cromwellian (@cromwellian) reported 17 minutes ago

  @perrymetzger @NathanLeamerDC GitHub went down for a few hours, and literally the entire tech development sector shut down.

• 
  Rituraj (@RituWithAI) reported 18 minutes ago

  🚨BREAKING: GitHub's AI security tool just created the vulnerability it was supposed to prevent. And attackers exploited it within five days. Wiz Red Agent found that a GitHub Copilot Autofix patch to Snowflake's snowflake-connector-net repo on June 18, 2026 replaced a safe input pattern with raw string interpolation of a GitHub issue title, opening a shell-injection hole exploited within five days. Read that again. GitHub Copilot Autofix is the AI tool that scans your code, finds security vulnerabilities, and automatically patches them. It's marketed as making your codebase safer. It's built by the company that runs the world's largest code repository. It introduced a shell injection vulnerability into Snowflake's production codebase. Attackers found it. Exploited it. In five days. The security tool was the attack surface. Here's exactly what happened. Snowflake's snowflake-connector-net repo had a vulnerable pattern. Copilot Autofix detected it and generated an automated patch — the kind of one-click fix that millions of developers trust every day because it comes from GitHub's own AI security system. The patch replaced the vulnerable pattern. But in doing so, it introduced raw string interpolation of a GitHub issue title directly into a shell command. Any attacker who could control the title of a GitHub issue could now inject arbitrary shell commands into Snowflake's build pipeline. A conditional gate that checked pull_request.user.login evaluated true on issue events because pull_request was null, letting an unauthenticated attacker through. Unauthenticated. Anyone. No credentials required. The AI generated the fix. The fix was reviewed. The fix was merged. The fix was the vulnerability. Here's why this is more alarming than a standard security breach. When a human developer introduces a security vulnerability, it's a mistake. An oversight. A gap in knowledge or attention. When an AI security tool introduces a vulnerability, it's a systemic failure. Every developer who trusted Copilot Autofix's suggested patches — across every repo it touched — now has to ask: did it help or did it hurt? The entire value proposition of AI security tools is that they're more consistent than humans. They don't get tired. They don't miss things. They catch what developers overlook. Copilot Autofix didn't overlook a vulnerability. It created one. Then approved it. Then merged it into production. Here's the scale of what this means. GitHub Copilot Autofix is used across millions of repositories.Every automated security patch it has ever suggested is now under scrutiny. Not because developers should have caught it — but because the tool that was supposed to catch it was generating the problem. How many other Copilot Autofix patches introduced vulnerabilities that nobody has found yet? How many are sitting in production codebases right now, waiting for an attacker who bothers to look? Here's the brutal irony that nobody is saying out loud. This happened the same week GitHub Copilot switched to token-based billing. Developers are already angry about surprise charges for agentic sessions. Now the tool is also introducing shell injection vulnerabilities into their production code. You're paying more. For a tool that just compromised Snowflake. Here's what you should do right now. If your team uses GitHub Copilot Autofix — audit every automated patch it has merged in the last 90 days. Not just the ones that looked suspicious. All of them. The Snowflake patch looked clean. It passed review. It wasn't clean. The AI that was supposed to make your code safer just made it less safe. And it did it automatically. At scale. Without anyone noticing for five days.

• 
  Larry Chiang, 650-283-8008 (@LarryChiang) reported 19 minutes ago

  Was asked how I think about the things I work on holistically. I break them into 3 buckets: 1. Random productivity hacks or experiments. These are usually an evening or weekend vibe-coding session. If something ends up useful or packaged enough, I’ll push it to GitHub. 2. More in-depth products with no intention of making money. I use these as playgrounds for learning and experimenting with AI. They can easily turn into 100s of hours of focused product + AI work, but they’re a great way to test workflows and hopefully build something useful for the world. Examples are GPT Food Cam and Speakrphone. 3. Things that can make money. These tend to be more deeply researched problems, often in a specific vertical, where there’s a clear user, buyer, and willingness to pay. A lot of what I build at work falls into this bucket.

• 
  Charlie Barmore, CPA, CFE, CVA (@cbarmorecpa) reported 22 minutes ago

  I’ve gotten a lot of positive feedback since open-sourcing LedgerTB. Even better, I’ve started working directly with other accountants who are using the software and helping improve it. One of those accountants is @Scotchua. Scott ran into a bug when using the app, worked through the problem with Claude, figured out a fix, and Claude told him, “This would make a good pull request.” So he opened one on GitHub. We went back and forth through comments and reviews. I could have just made the change myself, but I wanted us to go through the full process, and I wanted him to get proper credit for the contribution. Since then, we’ve continued trading ideas, fixes, and tests. Scott is using his agent. I’m using mine. He even realized Claude could test the app on Windows through a Parallels installation he was about to delete. So now we’re bringing two accountants, two agents, two environments, and two different sets of assumptions to the work. He catches things I miss. I catch things he misses. We review each other’s changes and try to make sure fixing one thing doesn’t quietly break another. That back-and-forth has been one of my favorite parts of open-sourcing LedgerTB. Neither of us have an engineering background. AI is allowing us to participate much more directly in the development process, but that doesn’t make software engineering expertise any less important. Accountants know how the workflow should operate, what the output should mean, and where the accounting risks are. Engineers bring a different kind of judgment around architecture, security, reliability, and maintainability. The interesting part is that AI is making it easier for those perspectives to meet. It’s shortening the distance between “I wish the software did this” and “I’ve built a proposed change, tested it, and submitted it for review.” That doesn’t eliminate the need for engineering rigor, but it does give domain experts a much more active role in the process. I think that’s one of the biggest opportunities in open-source, AI-assisted software: the people who understand the problem can work more directly with the people and the tools that understand how to build reliable solutions. And it’s a lot more fun when we build them together!

• 
  Hari (@harii_07) reported 29 minutes ago

  @JamesNguyen868 @thekitze That's an L take. You can't complain about github being down if you haven't built github before? You can't report an issue if you dont work on a competitor?

• 
  Chris (@chrisveleris) reported 31 minutes ago

  tududi passed 3200 GitHub stars this week. What moved it: - a 20 second demo gif at the top of the README - a docker compose that works on the first run - replying to every issue within a day and fixing bugs instantly

• 
  Brooks Chambers (@BrooksChambers) reported 33 minutes ago

  @adrianmg sorry it’s like a github issue for someone who occasionally touches grass but doesn’t like it

• 
  Shreyas Nalle (@ShreyasNalle) reported 33 minutes ago

  Were they really waiting for github to go down for their launch Anyways, competing with github is not a joke, there are quite many alternatives for github but I beat many people havent even heard about it Lets see how this goes