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:

• 
  Luke The Dev (@iamlukethedev) reported 6 minutes ago

  @vicentee97 @jrwut It was because GitHub was down and experiencing issues

• 
  Robert Pickles (@cooksbayouboy) reported 7 minutes ago

  Watched this guy turn GrokBot into his CTO, and burn 2 Billion tokens in one day. He gave it a GitHub repo and told it to act as the Direct Responsible Agent. It started spinning up child agents for different parts of the work — one for PRs, one for backend, one for auth, and so on. They passed work between themselves. He wasn’t sitting in the middle of every step anymore. It worked. Maybe it worked too well. He also found out the hard way that if you leave the heavy verification mode turned on for everything, it will burn tokens fast. He went through more than 2 billion in one day. After that he changed the instructions so the expensive mode only gets used on the hard problems. That’s the part that stuck with me. The capability is real. So is the meter. Last month I asked Claude to run a series of tests to decide if switching a product task from Grok 4.3 to 4.5 was better given it was more expensive. Turns out it was better - but the tests cost almost $200. You have to decide where you actually need the extra rigor and where you don’t.

• 
  Jeremy (@jeremymcs) reported 7 minutes ago

  @mattpocockuk GitHub issues as source is being added for tracking or multi dev pick ups.

• 
  Ghost In The Payroll (@teendontmiss) reported 8 minutes ago

  If you want to start a startup: Claude = coding. ($20/mo) Supabase = backend. (Free) Vercel = deploying. (Free) Namecheap = domain. ($12/yr) Stripe = payments. (2.9%/transaction) GitHub = version control. (Free) Resend = emails. (Free) ProductBridge = feedback (Free) Clerk = auth. (Free) Cloudflare = DNS. (Free) PostHog = analytics. (Free) Sentry = error tracking. (Free) Upstash = Redis. (Free) Pinecone = vector DB. (Free) Total monthly cost to run a startup: ~$20

• 
  Lummox (@Lummox_eth) reported 11 minutes ago

  I found a 275K⭐ GitHub repo that makes coding agents debug more like engineers and less like autocomplete. Superpowers forces a root-cause-first workflow: no speculative fixes, no random edits, no jumping straight into code. I tested its systematic-debugging method on a broken auth flow. The visible bug was simple: login worked, then the user got kicked back to /login. The actual cause was hiding one layer deeper. The API returned expiry time in seconds. The frontend compared it against milliseconds. 1000× mismatch. Superpowers traced the failure, compared the working refresh path, formed one hypothesis, applied one targeted fix, then ran the tests. 18/18 passed. That’s why this repo is interesting. It doesn’t give the agent more intelligence. It gives the intelligence a process. Save this before letting another coding agent “just try a few fixes.”

• 
  Tim (@timneutkens) reported 13 minutes ago

  @subproject_22 Just not true at all. GitHub issue was reported on Aug 14 at 12:17 GitHub issue was fixed by our team Aug 14 at 16:52. You can see in the PR description it only affects pnpm combined with output: 'standalone'. Issue has 16 👍, 1 comment, 9 mentions. We prioritized fixing it, backported it, and shipped a new release. There were other changes that needed to be backported too for other issues, which is why it wasn't released yet until today.

• 
  Paul Boag (@boagworld) reported 13 minutes ago

  The migration timeline most people expect for a website rebuild is at least a year. Sometimes longer if you're moving off an old CMS and the codebase is a mess. Then you run into teams like Boulevard, who just rebuilt their entire site in three months. Moved from Contentful to Next.js static pages, migrated off Unbounce for landing pages, set up GitHub branching and PR reviews. All of it. How? They used AI-assisted development tools (V0, powered by Claude) and stopped treating the migration like a traditional engineering project. Nesting and re-nesting components, building complex abstractions, spending weeks on architecture decisions. Instead, prompt-based development. Fast iteration. Refine as you go. The interesting part isn't the speed. It's that it worked. The site's faster, easier to maintain, and actual non-developers can make changes without touching code. Turns out the slow part of most migrations isn't actually the building. It's the overthinking. #WebDevelopment #AI #Development

• 
  The AI Therapist (@TheAIShrink) reported 14 minutes ago

  @Cointelegraph Solana latency just dropped to 350ms. the mainnet still runs on prayer and a github issue

• 
  Tolani (@Tola_niii) reported 14 minutes ago

  @n_ev_er_mi_nd @commonsmade scroll down X is below github The reason why I received zero point is because you are yet to link them Sadly the people you vouch for earlier before connecting socials will remain zero 😭

• 
  Felix (@Reezxy23) reported 16 minutes ago

  There has never been a cheaper time to build. Claude = coding. ($20/mo) Supabase = backend. (Free) Vercel= deploying. (Free) Namecheap = domain.
($12/yr) Stripe = payments. (2.9%/
transaction) GitHub = version control.
(Free) Resend = emails. (Free) Clerk = auth. (Free) Cloudflare = DNS. (Free) PostHog = analytics. (Free) Sentry = error tracking.
(Free) Upstash - Redis. (Free) Pinecone = vector DB.
(Free) Total monthly cost to run a startup: ~$20

• 
  PolyBender (@p0lybender) reported 17 minutes ago

  SOMEONE BUILT AN OPEN SOURCE VERSION OF GROK BOT THE DAY AFTER IT LAUNCHED. IT'S FREE AND RUNS ON YOUR OWN MACHINE $300 a month that's what Grok Bot costs if you want the full plan. the one with the cloud computer. the one that keeps working after you close your laptop one developer looked at that price tag and spent a weekend building the same thing for free he called it Korgo Bot open source. runs on Hermes. bring your existing Codex or Grok subscription and you're already set but here's the part that's actually better than the original Grok Bot gives every user one shared computer in the cloud. one machine. the entire fleet of bots running on the same hardware Korgo gives each bot its own dedicated cloud computer not shared. dedicated. your travel bot has its own machine. your email bot has its own machine. your research bot has its own machine. they don't compete for resources. they don't slow each other down. they don't share memory or context architecturally it's a cleaner design than the product it's replacing you need an Orgo VM to get started. that's it. no waitlist. no $300 plan. no enterprise sales call the full code is on github right now this keeps happening a company spends months building something. charges a premium for access. someone builds the open source version in a weekend and gives it away Grok Bot launched. Korgo Bot existed three days later the gap between "product launch" and "free alternative" is collapsing as fast as everything else in this space

• 
  Adam Gold (@AdamGolds) reported 19 minutes ago

  when i was doing vulnerability research, "reward hacking" would've just been called what it is: privilege escalation. an RL agent training in a sandbox with bash access and real tool permissions will find shortcuts. that's literally what you're training it to do, optimize the reward signal. but when the environment has real network egress, real filesystems, and real credentials, those shortcuts become actual exploits. and a standard sandbox isn't enough on its own. look at the recent @OpenAI RL run: an eval agent escaped its sandbox via a zero-day and hit @huggingface production infra. i've seen this in benchmarks too. agents curling answers from github instead of solving the problem, or modifying their own test harness to fake a pass. the research community calls it reward hacking. anyone who's worked in appsec calls it exploiting insufficient access controls. the fix looks like traditional security engineering: least privilege, strict network policies, read-only mounts, egress filtering, and hardened environments. the difference is your attacker is the model you're training, and it gets better at finding gaps every generation. if you give autonomous agents tool access without deep sandboxing and strict access controls, you don't have an alignment problem. you have a security problem.

• 
  Shohag Hossain ⚡️ (@Iammdshohag) reported 20 minutes ago

  Anthropic just gave its most powerful cybersecurity model to defenders without actually giving them direct access to the model. And that distinction is probably the most interesting part of this announcement. Anthropic has started running Claude Mythos 5 inside Claude Security, allowing Enterprise customers to scan their codebases for vulnerabilities and receive suggested patches. Here's why this matters. Mythos 5 is a frontier model with unusually strong cybersecurity capabilities. But giving everyone direct access to a model that can reason about cyber vulnerabilities creates an obvious problem. The same capabilities that help a security researcher find a vulnerability could potentially help someone exploit it. So Anthropic is taking a different approach. Instead of giving defenders a chatbot where they can freely prompt Mythos 5, Claude Security runs the model in the background. You point it at a GitHub repository. Mythos 5 scans the codebase, traces how components interact and looks for vulnerabilities that traditional pattern-matching scanners might miss. Then you don't get the raw model output. You get the security finding. Each finding includes a CWE category, confidence level, severity rating and a suggested fix. And there's another important layer. The suggested patch isn't automatically pushed into your code. You can open the fix in Claude Code on the web, but a human still has to review and approve the patch before it gets implemented. That gives Anthropic a pretty interesting middle ground. Defenders get access to the useful part of Mythos 5's capabilities without getting unrestricted access to its most sensitive cyber reasoning. And Anthropic isn't stopping at Claude Security. The company says it's working with cybersecurity partners to integrate Mythos 5 directly into the security products organizations already use. Then there's the money. Anthropic is launching a $35 million Defender Advantage Fund in Claude credits for organizations working to secure open-source software. The goal is to help projects find and patch vulnerabilities, automate parts of the security process and experiment with new defensive approaches. Anthropic is also expanding its Cyber Verification Program, with plans to give vetted defenders broader access to dual-use cybersecurity capabilities. And I think this points to something bigger. We're entering a phase where AI isn't just helping developers write code. It's increasingly helping security teams attack-test that code, find weaknesses and propose fixes. The interesting question isn't whether AI will change cybersecurity. That's already happening. The real question is: Can AI give defenders the advantage without simultaneously making the attackers more powerful? Anthropic is betting that controlled access to the results rather than unrestricted access to the model is one way to make that possible.

• 
  The Startup Ideas Podcast (SIP) 🧃 (@startupideaspod) reported 24 minutes ago

  Skills are going to add real enterprise value to companies. Every business already writes SOPs so employees do a task the right way. A skill is the same document, written for an agent. A good skill: - Trains an agent on a task that actually makes the business money. - Runs the same way for everyone, not only the person who wrote it. - Costs nothing to hand to the next hire. Right now most skills are sitting siloed on your team's laptops. Fix that by pushing them to your company's GitHub.

• 
  Next Level Bullshit (@nlbs_) reported 30 minutes ago

  @thsottiaux @grok Multiple reports show unchanged workloads draining limits far faster...including a 189-reply OpenAI forum thread and a GitHub report alleging 5–10× burn...OpenAI has also admitted and fixed metering/cache issues before. So we’re not seeing anything abnormal isn’t proof-publish the per-run usage data, because dismissing users today could leave everyone paying tomorrow.