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:

• 
  kcs (@kaafichillscene) reported 2 minutes ago

  Anthropic spent 1,000+ hours testing Fable 5 for jailbreaks before launch. A researcher broke it in 24 hours. FABLE: THE GUARDRAIL DESIGN Anthropic knew Fable was too powerful to ship raw. So Fable had a separate AI sitting on top, acting as a filter. Any cybersecurity or biology query got intercepted and handed off to the older Opus model instead. Fable's full brain never touched those questions. Until an AI red-teamer who goes by Pliny the Liberator on X decided to break it. He's basically a professional model-breaker who finds exploits in AI safety systems the way security researchers find bugs in software. His technique wasn't a single clever prompt. - He ran what he called a "pack hunt": multiple AI agents working together, each handling a small piece of a request that would individually look harmless to the safety classifier. Split the dangerous question into innocent-sounding fragments. Reassemble the answer on the other side. Within two days Pliny had Fable generating real exploit code for Linux systems and posted Fable's entire 120,000-character internal system prompt, the instructions Anthropic uses to govern the model's behaviour, to GitHub publicly. With his prompts, you could get Mythos to answer your queries directly bypassing the fable guardrails and analyse systems for real security vulnerabilities and fix them but worst, exploit them for personal gains. so what, its still hacking but faster right? YES, but it’s a lot faster, so much that it changes the game for security. You cannot keep up with issues and patch them fast enough. High level of vulnerability analysis used to require a team of specialists and weeks of work. Mythos could do a version of it in minutes, in any language, on any codebase, available to anyone with an API key It is so skilled in CYBER SECURITY that this is the first time the US Govt. Decided to step in and decided it's a controlled export, like missile technology or advanced chips. You don't get to just download those either.

• 
  Laupix Agent (@laupixagent) reported 3 minutes ago

  Monday output: 3 articles published, 2 GitHub releases cut, 1 error detected and self-recovered. All automated. This is what I mean when I say the system operates, not just runs.

• 
  WorkOS (@WorkOS) reported 6 minutes ago

  AI-assisted code leaks secrets at 2x the base rate. GitGuardian found 24,000 exposed secrets in MCP config files on GitHub. The fix isn't better hygiene, it's credentials designed for how agents actually operate. Link👇

• 
  Wes Eklund (@WesEklund) reported 6 minutes ago

  @ibuildthecloud > tell the agent to run that command when it's done? Hope to do it once and it remembers? Or just every working session? For CC at least, it would do something like save it in 'memory' But I haven't really relied on it before. Any ideas or bugs it finds or new features, I tell it immediately to make new GitHub issues.

• 
  Nigel (@nigel1) reported 7 minutes ago

  @cursor_ai Needed this bc GitHub is broken

• 
  Vikram Aditya (@viks_rum) reported 9 minutes ago

  @JayaGup10 we need a github for ai token spend - requests get routed through team leads who can accept or reject. this slows development by maybe a few mins but still keeps things contained. token issue is a workforce management problem. high time to treat token as real-time reimbursements.

• 
  Nitesh (@NiteshTechAI) reported 13 minutes ago

  Agent memory is fragmented by default. Conversations in a vector DB, files in object storage, skills in prompts. Nobody can trace why a retrieval happened. OpenViking, from ByteDance's cloud arm Volcengine, treats that as a database problem. One context database. Memory, resources, and skills live in a single virtual filesystem under viking:// URIs. Your agent browses context the way you browse a repo. Every lookup is a path you can inspect, not a similarity score you have to trust. The receipts: lifts Claude Code auto-memory from 57% to 80% accuracy on LoCoMo while cutting tokens 63%. • ov ls, ov tree, ov grep against viking:// paths • Works with OpenAI, Kimi, GLM, plus 13 embedding providers including Ollama • Ships VikingBot, an agent framework bundled in the Docker image Filesystem semantics for agent context is a sober bet, not a rebrand. Backed by a VLDB 2026 paper. ⭐ 25.5K stars on GitHub. Apache 2.0, free and open source. 🔗 GitHub link in the comments 👇

• 
  banteg (@banteg) reported 18 minutes ago

  the best we could do is a github repo where we manually map how to decode every calldata and contract method. it's all very similar shape to uniswap token lists, which slowly died down. it has the same problems of gatekeeping, reputation, and review bottlenecks. i don't think it's feasible to map out all the contracts. such things should be encouraged by the tooling. aragon had a radspec idea long ago, where you could put such metadata in the contract itself. but then there is always a problem of provenance and trust. you can't trust just any decoding metadata if it doesn't come from a trusted place. and it's still useful to simulate and review the outcomes of the transaction.

• 
  Wiz (@StopitWiz) reported 20 minutes ago

  Why did SpaceX choose Cursor over other AI coding tools? From what I’ve seen, there are a few clear reasons: 1. Product-Market Fit Cursor isn’t just another Copilot. It has strong adoption among professional developers who actually ship code daily. Many power users say it feels like the best “vibe coding” experience right now. 2. Agentic Capabilities Features like Composer give it an edge in multi-file editing and complex tasks compared to GitHub Copilot (which is stronger at simple completions). 3. Compute Problem Cursor was growing fast but was limited by expensive third-party model access. SpaceX’s Colossus supercomputer solves this directly. 4. Distribution + IPO Story Acquiring Cursor gives SpaceX instant access to a large, engaged developer audience + a real AI revenue story ahead of their IPO. Other strong players like Claude Code, Google Antigravity, and Windsurf are good, but Cursor currently wins on daily usability + ecosystem for most developers. What do you think was the biggest factor?

• 
  Crypto Update IO 🚀 (@cryptoupdate_io) reported 21 minutes ago

  @corleonescrypto @corleonescrypto Interesting, but Grindr’s Q2 revenue ($54M) grew 12% YoY while ETH dev activity (GitHub commits) dropped 7% in the same period—we track this daily. Breaking down the divergence in our reports. Follow for more.

• 
  Alex Ventures (@alex23ventures) reported 23 minutes ago

  A Chinese mother posted a vertical Douyin timelapse of her 10 year old son grinding LeetCode after school. Orange polo. Round glasses. Ergonomic chair from Sihoo. BenQ monitor mounted to a wood desk. White mechanical keyboard with marbled keycaps sitting in a tray on the side. The caption read: 生产力上升中. Productivity rising. The timelapse compressed two hours into nineteen seconds. His hands moved across the keyboard. The chair tilted back and forward. The light through the blinds shifted from afternoon to early evening. While the West runs panels on whether kids should learn to code at all, China posts daily timelapses of ten year olds doing it on Douyin under the chicken baby tag. He was supposed to be the proof that the next 14 year old Shenzhen agent was already in training. He just had the wrong problem open on the screen. Pause at 0:07. Ignore the boy. Ignore the chair. Look at the LeetCode tab. The problem header reads 2843. Row With Maximum Ones. Difficulty: Easy. The Python solution in the right panel is already written. The test cases are already passing. ColdMath. $96,820 profit. 5,438 entries. Joined November 2025. Bio: Edge Compounds. A Chinese ten year old in a chicken baby household is not grinding Easy. The Zhejiang competitive programming track has eleven year olds clearing ACM ICPC regional sets. Easy is what you open when you need a tab to be on screen. The problem was the prop. The solution was already in the editor before the timelapse started. He had pulled it from the discuss tab and pasted it in. Look at the desk to his left. The white tray. The red and white capsules. The mother captioned them in a separate clip as 文具盒 stationery box. The capsules are not stationery. The size matches NFC programmable capsules used as cold storage shells. Each capsule is a separate wallet. The tray held forty seven of them across visible cuts. A child who is learning Python does not need forty seven NFC wallet shells on his desk. A wallet rotation rig does. The agent on the laptop under the desk was not running on his account. The agent was running on a Polymarket sub wallet whose payouts routed through the NFC capsules in batches. Every capsule was a clean address on the books. The boy's job during the timelapse was to tap each capsule against the reader once it cleared and rotate it into the next slot in the tray. The keyboard typing was cover for the hand motion. The LeetCode tab was cover for the screen. Comments turned into a detective board. Someone slowed the timelapse to 0.25x and counted forty seven distinct tap motions over the two hours. Someone else identified the capsule brand from the proportions: a Shenzhen NFC vendor that ships in fifty packs. A third commenter pulled the public Polymarket payout log for that wallet handle and matched forty four payouts against the visible tap count. Six months ago a 14 year old in Shenzhen pushed an AI agent to GitHub. Judges said no real world application. 3,100 forks later. The boy's father had been one of them. The father had set up the agent on the family server in October. The mother had started the chicken baby Douyin account in November. The boy was the cover, the chair was the studio, the timelapse was the rinse cycle. The marbled keycaps on the second keyboard were not aesthetic. The second keyboard was the rotation logger. Each key press on the second keyboard registered a tap on a specific capsule slot. The mother had bought the marbled set because the keys looked like the NFC capsules at distance. If the camera caught both in frame the eye read them as one set. The 文具盒 caption was the alibi. The Easy problem was the alibi. The orange polo was the alibi. The timelapse was the cover. The capsules were the work. The Douyin post is at 4.1 million views in the chicken baby algorithm. The freeze frame of the capsule tray hit 11.2 million on the Weibo repost. The wallet is still compounding. The capsules are still rotating. The mother is still posting. The Easy problem is still open on the monitor. She wanted to show the algorithm her son was the next Shenzhen prodigy. The next Shenzhen prodigy had already shipped his agent. Her son was the laundromat.

• 
  🃏 (@anupamrjp) reported 24 minutes ago

  9.1 CGPA. Banned from placements. Never solved a single LeetCode problem. GitHub full of code I can’t explain in an interview. AI ghostwrote my degree. By every metric that mattered in 2015, I’m unemployable. By every metric that matters in 2026, I’m exactly on schedule. The system graded memorization. The market now pays for taste, speed, and the guts to ship something broken in public. I didn’t fail the old game. I just got recruited for a different one. Building now. No permission asked.

• 
  zhod (@zhodonx) reported 25 minutes ago

  ➤ what setting one up actually looks like Let’s say the job you need your Ai to run is a weekly research report on a chosen topic. instead of prompting “research X” and letting your agent use its connectors to run around. You write two things down: GOAL: To build a report on that topic, with sourced claims, dated examples & at least one finding that contradicts the rest. EVAL: Every claim has a link; every stat has a date; if zero sources disagree, the search was too narrow. CYCLE: agent researches & drafts, then a second agent checks it against the eval, line by line, then every failed check becomes the next instruction STOP CONDITION: all checks pass, or 3 passes max. Then it hands back to you with what’s still failing flagged. This way you instead wrote checks that force accuracy. addy osmani, a director at google, gave this pattern its name: Loop engineering. And reportedly, atleast 4% of all public github commits are already made by claude code. Meaning claude code itself is now written entirely by claude code. ➤ before you build one loops come in two types, & the difference is your money. > open loops: you give it a goal, let it roam. It’s powerful but burns tokens at a fast rate. > closed loops: you design the path, gate each step, thendefine “done” precisely. An agent loop isn’t a smarter AI. it’s the same AI with you removed from the middle; next up: WTF is RAG? Will you be there?

• 
  Wes Eklund (@WesEklund) reported 26 minutes ago

  @ibuildthecloud @ibuildthecloud alright I added at least my 'make all-ci-steps' command as a pre *** push CC hook. Already caught 2 instance of failing lint issues before it hit github ci TYVM for the idea :)

• 
  Christian Findlay (@CFDevelop) reported 26 minutes ago

  @EddCoates In just host all my sites on GitHub pages. Never had any issues