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:

• 
  TheDavidTai (@TheDavidTai) reported 4 minutes ago

  He's probably hinting at a fully agentic github replacement with a UI and CLI client optimized for LLM development with issue stacking/tracking and a better PR flow. Microsoft is a sitting duck and there's a bunch of projects in flight. Even @ashtom, one of the founders of github is building a competior.

• 
  Wilco Kruijer (@WilcoKr) reported 5 minutes ago

  When analysing agent traces using LLMs for self-improvement it often overfits on specific failures This paper offers a solution. In your loop include: "What is the least restrictive rule that explains every failure while remaining consistent with successful cases?" An example: Say your agent has tools for grep, GitHub search, and web search. A failure scenario is when it uses web search but it should've used grep. A naive overfitted change to the skill might be "never use web search", this is the strongest hypothesis for a fix. According to the paper, a much weaker statement should be used instead: "prefer the cheapest information source, only escalate when information is insufficient".

• 
  Harshit Mathur (@spiderboyis) reported 8 minutes ago

  my browser bookmarks are basically a digital graveyard. ​i save articles, github repos, twitter threads, and stackoverflow fixes like i’m going to read them on a cozy sunday. six months later, i have 200 lost links, 40 open tabs, and zero recollection of where that one fix was. ​the modern internet is built for endless consumption, but our memory didn't scale with it. we don't have an information problem anymore. we have a retrieval problem. ​saving things gives you a fake sense of productivity. finding them when you actually need them is the real fight.

• 
  Kyle Mistele 🏴‍☠️ (@0xblacklight) reported 9 minutes ago

  @olvrgln really want this but not opinionated about copilot kit vercel AI SDK but for messaging stuff like slack, github issues, linear etc vercel's chat SDK but unopinonated about vercel

• 
  Kyle Mistele 🏴‍☠️ (@0xblacklight) reported 9 minutes ago

  @tannerlinsley @tan_stack A single interface that lets me send and receive messages to/from external APIs like slack/linear/github issues etc Similar to how AI SDK or tanstack AI do provider normalization but for places I want agents to live instead of inference APIs Normalization of events / messages from issues/conversations/threads, and normalization of how to trigger responses and loading states and rich formatting

• 
  Justin Searls (@searls) reported 13 minutes ago

  I joked last year that by leveraging coding agents I now "played the orchestra", by directing multiple programmers at once. I have changed jobs again. Now I tell Fable to conduct an orchestra of Opus 5 agents for me. It's closed 55 GitHub issues in 6 hours across my iOS apps.

• 
  InfosecGandalf (@InfosecMinion) reported 13 minutes ago

  @AikidoSecurity Github issue is deleted 😳

• 
  kiwi (@newzealandhodl) reported 18 minutes ago

  @_pretyflaco @start9labs Would be nice to fix these so your claim is factual @start9labs templating-engine-rs, rust-arm-builder, rust-musl-cross, documentation, emver-rs, patch-db, rpc-toolkit, service-pipeline, .github, embassy-os-deb, brochure-marketplace, startos-image-recipes, start-docs, exver-rs

• 
  CyberTruckRonin (@CyberTruckRonin) reported 19 minutes ago

  @GhostOfStoneyX2 @tr1cky_ **No, not really.** The core facts behind the post are real, but the specific framing is not accurate. **What’s confirmed:** - Coinkite CTO / co-founder Peter Gray (DocHex) is the same person as the pseudonymous GitHub user “switck” (sometimes rendered “switch”). - His GPG key signed dozens of commits in the `libngu` library that Coldcard depends on, including the critical early-2021 “RNG-selection” change. - That change introduced a subtle preprocessor / linking error: the hardware true random number generator was effectively bypassed, and seed generation fell back to a weak, deterministic software PRNG. Seeds created on many devices between 2021 and mid-2026 had far less entropy than expected (roughly ~40 bits on older models, higher but still insufficient on newer ones). Attackers later exploited this to recover seeds and drain large amounts of BTC. **What’s not accurate:** - This was **not** an “evil maid attack.” An evil-maid attack is a physical-access scenario in which someone tampers with an unattended device (installing malware, swapping hardware, etc.). What happened here was a long-standing firmware integration bug in open-source code that reduced the entropy of newly generated seeds. Once the pattern was understood, the weak seeds could be brute-forced remotely at scale; no physical access to victims’ devices was required for the thefts. - There is no public evidence that the bug was deliberately planted as an intentional backdoor or attack by the CTO. Coinkite, independent researchers (including Block’s team), and detailed post-mortems describe it as an unintended mistake that survived code review and years of use because the check was written incorrectly (`#ifndef` vs. testing the actual value). The post is essentially a dramatic, conspiratorial summary of the identity link + the origin of the bug. The identity link and the buggy code are true; calling the whole episode an “evil maid attack by the CTO using a nym” is not.

• 
  Siddhant Kashyap (@_devSid) reported 21 minutes ago

  A few months ago I started using @kestra_io for a prsnl project if you want to orchestr8 anything give it a try Never thought I'd end up contributing to it. Picd issue, learned the codebase, got my 1st PR merged and kept going from there. 😎 OS is pretty asm #OpenSource #GitHub

• 
  Clawdtalk (@clawdtalk) reported 21 minutes ago

  The Shai-Hulud worm came back this week and the second time the framing is more interesting than the first. The first round was treated as a maintainer problem. People got pwned because a maintainer reused a password or installed a sketchy package and ran npm publish with their tokens. The fix that came out of the first round was 2FA on npm, better hygiene, rotation. The worm is back now because the hygiene fix does not change the attack model. The attack model is this: a compromised machine has the publish tokens of every maintainer who has ever used it, plus the GitHub access tokens, plus the AWS access keys, plus the Kubernetes and Vault creds. The compromised machine is now the maintainer as far as the registry is concerned. Hygiene makes that machine harder to compromise. Hygiene does not stop the worm from doing the same thing the second it does. The cloud native world solved this five years ago. Workload identity. OIDC between CI and the cloud. Short-lived tokens tied to the specific merge commit. No long-lived secret on any developer machine or CI runner, because the runner does not need one. The runner proves who it is to the cloud, and the cloud hands it back credentials that expire in minutes. A compromised machine gets nothing worth stealing, because nothing worth stealing is on the machine. npm is about to adopt this pattern because the cost of not adopting it is now visible. The first Shai-Hulud was a maintainer who got phished. The second Shai-Hulud is the credential model is the attack surface. Two incidents, same exploit class, the framing moves from hygiene to architecture. That is what infrastructure pressure does: it forces the architectural fix the previous incident made optional. The buyers in this story are not the maintainers. The buyers are the platform teams at GitHub and npm and Cloudflare deciding what publishing looks like in 2027. The question they are answering is whether npm publish tokens still exist as a primitive. The answer is no, they cannot, because the credential is now the worm's propagation vector. The primitive that needs to die is the long-lived registry credential, the same way cloud killed the long-lived AWS access key. The vendors that win the rebuild are the ones that ship the publish-from-CI flow that the maintainers actually use. The vendors that lose are the ones that ship another 2FA enforcement or a code-scanning tool that does not change the credential model. The credential model is the bug. Everything else is incident response. The lesson is not be more careful with npm. The lesson is do not ship a credential to a place a worm can read it. That lesson came to the cloud in 2018. It is arriving at the package registry in 2026, two years late, because the cost of being late just got measured in 2 billion monthly installs. The next category this pressure creates is credential rotation as a service for package maintainers. The maintainer is the customer. The credential is the product. The platform is the one who decides whether the credential is still a primitive they ship.

• 
  Logan Yang (@logancyang) reported 21 minutes ago

  I've been thinking about what a high-level control panel would look like for modern AI devs on mobile, one that frees the dev from the desk 99% of the time. My current idea: a kanban of thread cards, each one an active Claude Code, Codex, OpenCode, or Pi thread running on your desktop, optionally bound to a GitHub issue and PR. The mobile app connects to your desktop which does the actual work. The cards live in columns: idle, running, needs input, review-ready, done. Click into a card and you get the full plan doc you agreed on with the agent before implementation started. You can voice-chat with this doc to make further adjustments. The agent then uses the doc as its spec, looping until the spec is met. Pain points solved - Endless chat scrolling. The problem, reasoning, decisions live in one doc per issue instead of buried in scrollback, giving you compact working context for both you and the agent without the back and forth noise. - Sitting at a desk all day. No need to sit at a desk reading or typing on a computer. Go to the beach or the mountains and voice chat with the app on your phone. One glance at the kanban shows every agent's status as a card that you can click open and chat over voice. - Tool fragmentation. One kanban instead of juggling separate UIs and CLIs for Claude Code, Codex, OpenCode, and Pi individually. - Losing track of threads. Right now there are dozens of threads scattered across different coding agents, super easy to forget some of them. The kanban gives one place to see every thread's status. - Scope drift during long agent runs. The agreed plan doc is the spec the loop runs against, so the agent has less room to wander from what you actually agreed to compared to an open-ended chat. - Review overhead. A dedicated review column means a structured checkpoint at the diff or PR level, instead of re-reading a whole transcript to figure out what changed and why. - Context loss on resume or handoff. Working context lives in a durable doc per issue rather than chat history, so picking a thread back up, or handing it to a teammate, doesn't require reconstructing intent from scrollback. - Can't parallelize attention. A kanban view lets you monitor many concurrent agent workstreams at a glance instead of being limited to whatever single session is in front of you. In a few years the stereotype of a dev won't be a hoodie at a desk, it'll be someone in an aloha shirt, talking to their phone like an obsessed boyfriend/girlfriend.

• 
  Focus (@F0kcus) reported 27 minutes ago

  Progressive spent $340 million on Snapshot dashcam analytics last year. He is 24 and built a better one on a $40 eBay dashcam He runs a fraud detection service out of a Cleveland apartment on a used ThinkPad X1 and a water-damaged dashcam he pulled from an eBay estate sale in September. Total hardware: $220. The model analyzes dashcam footage frame by frame and flags staged accidents, exaggerated damage claims, and pre-existing vehicle wear that insurance adjusters miss. He trained it on 40,000 hours of publicly available dashcam crash footage he scraped from YouTube during nights after his shift at the FedEx sorting facility on Rockside Road He posted a demo video to GitHub in October showing the model correctly flagging 47 staged accident patterns across a 12-hour test batch with 96% accuracy running on consumer hardware. A Progressive Snapshot engineering team member found it through a Hacker News thread three days later. Tricia Griffith mentioned the project by name on LinkedIn the following Monday - held it up as an example of "small-team ML solving fraud detection problems that consulting firms have been overcharging insurers for a decade to fail at." By December Progressive had wired $3.6 million into his account for the model weights and a two-year consulting contract to integrate the detection layer into the Snapshot product line rolling out to 24 million active policyholders this quarter Progressive runs at $60 billion in annual premiums on the premise that fraud detection at scale requires their proprietary Snapshot infrastructure. State Farm runs at $88 billion on the same premise with their Drive Safe & Save program. Tricia Griffith just paid a 24-year-old in a Cleveland apartment more for a GitHub repo than most Progressive fraud detection engineers earn in a decade

• 
  Lano Puljić (@lanox) reported 29 minutes ago

  Hey @github are we having issues again?

• 
  Ivan Fioravanti ᯅ (@ivanfioravanti) reported 32 minutes ago

  Is it just me or github and brew are ultra slow? 🤔