GitHub Outage August 17, 2026: What Happened
The Short Answer
GitHub broke globally for about three and a half hours on August 17, 2026, and took Copilot down with it — for longer. Core services were mitigated by 16:59 UTC; Copilot authentication kept failing intermittently into the evening. GitHub has confirmed the fault was in its own service layer, not an AI model provider, and had not published a root cause analysis as of August 18.
Timeline (all times UTC, August 17, 2026)
| Time | Event |
|---|---|
| ~13:40 | Degradation begins; reports spike globally |
| 14:31 | Copilot enters “major outage” status |
| 14:31 | SAML/OIDC, SCIM and Team Sync added to affected components |
| 14:49–15:01 | Issues, Pull Requests, Actions and API Requests each enter major outage |
| 16:59 | Mitigated across API Requests, Actions, Git Operations, Issues, Pages, Pull Requests, Webhooks; moves to monitoring |
| 21:15 | GitHub still applying mitigations for “sporadic Copilot authentication failures in some applications” |
| — | Root cause analysis promised, not yet published |
What Actually Broke
The blast radius is the story. This was not a single feature failing — it was source control, CI, identity and AI assistance failing together:
- Git operations, PRs, Issues — the core version-control loop
- Actions — CI/CD pipelines stalled, blocking deploys
- API Requests and Webhooks — every integration downstream of GitHub
- SAML/OIDC, SCIM, Team Sync — enterprise SSO and provisioning
- Copilot — code completion, chat and agent workflows
The identity components are the underrated part. When SAML/OIDC degrades, organisations using GitHub as an SSO provider or SCIM source can see failures in systems that have nothing to do with GitHub itself.
The Copilot Detail That Matters
Copilot went into major outage before Issues, PRs, Actions and the API, and recovered after them. GitHub was explicit that the AI model providers behind Copilot stayed operational the entire time.
That points at GitHub’s own authentication and service layer as the failure domain — the plumbing that turns a Copilot request into an authorised model call. It is a useful reminder that an AI coding assistant is not just a model: it is a model plus token exchange, plus entitlement checks, plus routing. Any of those can fail while the model itself is perfectly healthy, and no amount of model redundancy helps.
Why This One Landed Harder Than Previous Outages
Three compounding factors:
1. Concentration. A team on GitHub + Actions + Copilot + GitHub SSO had four dependencies with one failure domain. In 2019 a GitHub outage stopped merges. In 2026 it stops merges, builds, logins and the assistant writing the code.
2. Agent workflows fail loudly. Autonomous agents that open PRs, push branches and poll CI don’t degrade gracefully during an API outage — they retry, burn tokens, and pile up failed runs that need manual cleanup afterwards.
3. The competitive timing. Cursor launched Origin, its own code hosting platform, the very same day. That was almost certainly coincidence — Origin had been visible in testing since mid-August — but it made the concentration argument for Cursor better than any marketing could.
The Correct Takeaway (and the Wrong One)
Wrong takeaway: migrate off GitHub. One incident, even a bad one, is not an availability trend, and every alternative has its own outage history. Migration costs weeks and buys you a different single point of failure.
Right takeaway: stop treating any single provider as infallible infrastructure. Git is distributed by design; a live second remote costs almost nothing and converts an outage from a full stop into a slow afternoon. Teams with a mirror kept committing and pushing on August 17. Teams without one watched a status page.
The practical playbook — second remotes, local model fallback, offline-capable tooling, and what to stop doing during an outage — is in how to keep shipping when your AI coding tools go down.
Last verified: August 18, 2026. GitHub’s root cause analysis was still pending at the time of writing; this page reflects status-page reporting.