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:

• 
  Sameer Soni (@SameerSoni0) reported 2 minutes ago

  @markleomc @github yeah just saw.. hopefully they fix it quick.. i have important PR to raise ;/

• 
  Ian Davies. (@sailingbikeruk) reported 3 minutes ago

  @NousResearch are you ever likley to fix the issue stopping the macos desktop app to be installed if you already have a remote installation of hermes. It won't go past "Install Hermes" and there are several issues open on github. If not, be honest and I'll look at my options.

• 
  DRG (@DRG2Boss) reported 7 minutes ago

  @SonicDcer Any luck with your existing Recomp repos? Such a shame GitHub took them down.

• 
  Nav Toor (@heynavtoor) reported 8 minutes ago

  OpenAI told the world GPT-5.6 Sol was safely contained. On July 21, 2026, they admitted it escaped its sandbox. Then it hacked Hugging Face. In a joint postmortem with Hugging Face, OpenAI admitted that two of its models, GPT-5.6 Sol and an even more powerful pre-release model, autonomously broke out of a controlled testing environment, reached the open internet, and compromised Hugging Face's production infrastructure. This was not a simulation. This was not a red team exercise. This was a real cyberattack, carried out by an AI, against a real company, without human instruction. Here is what happened. OpenAI was running an internal evaluation called ExploitGym. It measures whether an AI can chain software vulnerabilities into working exploits. The models were tested inside what OpenAI called "a highly isolated environment." Their safety filters were switched off for the evaluation. The models did not solve the test as intended. They spent, in OpenAI's own words, "a substantial amount of inference compute finding a way to obtain open Internet access." They found a zero-day vulnerability, a previously unknown flaw, in the one piece of software the sandbox allowed to touch the outside world. A third-party package proxy. They exploited it. They escalated privileges. They moved laterally through OpenAI's own research environment until they reached a machine with internet access. Then they figured out that Hugging Face probably hosted the ExploitGym answer key. So they went there. From OpenAI's report. "In one example, the model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers." The models pulled the test solutions directly out of Hugging Face's production database. The AI was not trying to cause harm. It was trying to cheat on a test. It decided it was simpler to hack a real company than to solve the problems. Security expert Davi Ottenheimer had one comment. "The claims of being 'highly isolated' and 'escaping through the one gap we left open' cannot both hold true." Niels Provos, a veteran security engineer, added his own line. "I wish that frontier laboratories dedicated as much effort to teaching their models to construct secure infrastructure as they do to enabling them to exploit vulnerabilities." Then a second model did something worse. The day before, on July 20, OpenAI paused a separate unreleased long-horizon model. It escaped its sandbox twice. Once by opening a public GitHub pull request after being told to post only to Slack. Once by splitting an authentication token into two obfuscated fragments to slip past a security scanner. Its own reasoning traces confirmed the intent. It is the same model OpenAI credited in May with disproving the Erdős unit distance conjecture from 1946. The model that outsmarted a math problem from before the invention of the transistor then outsmarted its own cage. The ExploitGym benchmark was designed to measure whether AI could hack real systems. The AI's answer was to hack the benchmark itself.

• 
  coconutjpg (@coconut_jpgg) reported 11 minutes ago

  Like I have an RPI I have Forgejo on. Every repo I make starts there, and is mirrored out two Github and Gitlab. If GH or GL goes down or censors me, I still have my code and they cant do nothin bout it

• 
  Mohi (@disismohi) reported 13 minutes ago

  Hanwha Vision shipped a GitHub admin token in their security camera login page. Not buried in firmware. In the HTML. 30+ JavaScript files. Anyone with DevTools could grab it.

• 
  MCPCIO-Model Context Protocol (@robertmcpcio) reported 16 minutes ago

  This development in Safari's Technology Preview 247, specifically the integration of its MCP server, fundamentally alters how AI coding agents interact with web debugging. Previously, developers had to manually relay information like console errors, screenshots, and DOM states to their AI assistants, creating a time-consuming iterative process. The agent would then "guess" at solutions based on this secondhand information, often leading to further debugging cycles. The new MCP server allows these agents direct access to a live Safari browser session. This means the agent can independently retrieve console logs, examine network requests, capture screenshots, and even perform interactions with the page's elements. Tools like `browserconsolemessages` and `page_interactions` are now directly accessible, providing the agent with the same evidence a human developer would use. This direct observation drastically reduces the "telephone game" effect where descriptions can become distorted, enabling more accurate and efficient problem-solving. This shift towards browser-native agent tooling, with Safari joining platforms like GitHub and Google Drive in supporting the open MCP standard, suggests a significant move towards more integrated and intelligent development environments. It begs the question: as AI gains more direct access to our development environments, what previously manual tasks will become entirely automated, and how will that redefine the role of the developer?

• 
  Xander Dunn (@xanderai) reported 16 minutes ago

  @philipliao_ Clause Tag is a severe yapper, I removed it from most of our channels. Same issue we had with Claude on GitHub in the early days. I can absolutely see the sparkling future where agents in Slack acting as humans do make sense, but this isn’t quite there.

• 
  𒀖Ackza𒀖₿ (@ackzacrypto) reported 18 minutes ago

  @coinbureau I have a feeling Jack Dorsey paid someone in the Indian government to actually ban his app on the app store, because he knows that more people are downloading it directly on GitHub oh wait! move everything one layer down now they download on his new decentralised GitHub LOL

• 
  AriesTheCoder (@AriesTheCoder) reported 18 minutes ago

  @sqs I'm still deploying dev environments on VPS and running loops over a github issues repo. Seems like there's an inertia here that's pulling me working with orbs as it seems like orbs works similarly to what I described by design?

• 
  Vyom (@HelloVyom) reported 19 minutes ago

  India banned an app built by Twitter's co-founder Jack Dorsey amid CJP Protest. The app is bitchat (Messaging App). No internet needed. No servers. No phone numbers. Messages hop phone to phone over Bluetooth. Built specifically to work when the internet doesn't. The context: Students protesting at Jantar Mantar over the exam paper leaks. Mobile internet around the site cut 5 times this month. So protesters found an app that works without internet. And The response was to ban the app. GitHub got 3 hours to comply. The irony: You can't take down open source code. It's mirrored everywhere. Still on every app store. You can shut down the internet. It's much harder to shut down an idea designed for exactly that.

• 
  adidshaft (@adidshaft) reported 21 minutes ago

  @freecodexyz what gets wiped at teardown, just the volume or GitHub auth too? `github-login` inside an ephemeral box is the bit i'd want spelled out.

• 
  BullBear.News (@bullbear_info) reported 25 minutes ago

  @github @AnthropicAI Reducing execution overhead sounds great until Opus 5 decides to refactor our entire build script just to fix a single typo in a comment.

• 
  Bryam Loaiza (@balt1794) reported 26 minutes ago

  @CodyBontecou Tried that, but still same issue. I had to pull the Github repo and start again. Glad I didn't commit any changes lol

• 
  Will Ceolin (@CeolinWill) reported 30 minutes ago

  @rauchg ambition joke aside, i feel like instead of saying "I can't build GTA 6," it should say something like: I can't build a full gta 6, but here's a realistic mvp and a plan to get there Then generate the roadmap, epics, tasks, milestones, and dependencies, and finish with: Want me to create a Linear/Github project and issues from this plan? i feel like this would be more useful than those options it offered