The cascade started with Microsoft, not Anthropic
On Thursday, November 12, Microsoft acknowledged issues reaching Exchange Online mailboxes across the Americas. By Friday morning, the outage had dragged into a second day, affecting Outlook, Teams, and the broader Microsoft 365 ecosystem. Government offices in Singapore, the Philippines, and Japan reported disruptions, along with India's IndiGo airline and parts of Britain's National Health Service. Australian officials compared the scope to the 2024 CrowdStrike failure. More than 130,000 users reported problems to Downdetector at Thursday's peak.
This was the substrate on which Friday's AI disruptions played out. Anthropic posted at 4:58 a.m. Eastern that it was investigating authentication issues affecting Claude.ai, Claude Code, and the developer console at platform.claude.com. By 6:05 a.m. Eastern, the company said it had deployed a fix and was monitoring recovery; Claude Code and the developer console returned to normal, but the main web app still showed a major outage.
Two distinct failure modes, one common dependency
What made this episode analytically useful is that Anthropic's status page surfaced two separate problems with different root causes:
- Authentication failures on Claude.ai, Claude Code, and platform.claude.com—Anthropic has not publicly linked these to Microsoft's outage.
- Microsoft 365 connector failures—tools that let Claude read Outlook mail, calendars, and files—explicitly attributed by Anthropic to "a Microsoft 365 service incident."
OpenAI reported the same connector breakage for ChatGPT's Microsoft integrations on Thursday. This symmetry matters: it confirms the failure domain was Microsoft's identity and graph APIs, not something specific to Anthropic's architecture.
Why the authentication outage matters more than it looks
The login problem on Claude.ai and Claude Code is the more consequential signal. Frontier model providers increasingly operate as identity providers for developer workflows: CLI tools, IDE extensions, and CI/CD pipelines all authenticate against Anthropic's or OpenAI's auth systems. When that auth layer wobbles, the blast radius extends beyond chat users to automated software delivery.
Anthropic's status page showed a major outage on claude.ai and partial outages on Claude Code and the developer console. The fact that Code and console recovered faster than the main web app suggests the authentication path for programmatic access (API keys, OAuth tokens for CLI) may be architected differently than the consumer web login flow—a detail worth watching in post-incident reporting.
The structural dependency nobody prices in
This incident illustrates a structural reality the market rarely prices: frontier AI labs are now deeply entangled with Microsoft's cloud. Anthropic runs significant training and inference workloads on Azure. Its Microsoft 365 connectors depend on Microsoft's Graph API and Exchange Online. OpenAI's ChatGPT connectors failed the same way. Even labs that diversify compute (Anthropic also uses AWS) still inherit Microsoft 365 as the de facto enterprise identity and data layer their customers demand.
The Economic Times noted that both Claude and ChatGPT reported service disruptions on November 12 and 13 "with users facing authentication problems and failures in Microsoft 365 integrations." The phrasing conflates two things, but the underlying point stands: the AI application layer has become a downstream consumer of Microsoft 365 reliability.
What to watch next
Three questions will determine whether this is a forgettable blip or a forcing function for architecture changes:
- Root cause disclosure: Will Anthropic publish a postmortem distinguishing the auth outage from the connector outage? The fact that they deployed a fix for authentication while Microsoft's outage continued suggests independent root causes—but confirmation matters.
- Connector resilience patterns: Will AI labs build fallback paths (cached credentials, degraded-mode operation, alternative identity providers) for Microsoft 365 integrations, or accept the coupling?
- Enterprise procurement pressure: Organizations that lived through the 2024 CrowdStrike incident and now this Microsoft 365 outage may start demanding contractual SLAs from AI vendors that account for upstream cloud dependencies.
Bottom line
Claude trended because a visible, user-facing AI service broke in two ways at once—one self-inflicted (auth), one inherited (Microsoft 365 connectors). The episode is a reminder that "AI reliability" is no longer just about model uptime; it's about the entire dependency chain from identity providers to enterprise data planes. The labs that acknowledge and architect for this reality will differentiate on trust. The ones that don't will keep trending for the wrong reasons.