Claude Code Auth Outage: claude.ai Login Failures and What to Do While Anthropic Recovers

Anthropic reported elevated authentication errors on claude.ai, Claude Code, and platform.claude.com starting November 13, 2026 at 09:58 UTC, with status.claude.com showing a major outage on claude.ai and partial outages on Claude Code and the console while users on Hacker News described login loops and hung sessions. If you depend on Claude Code for day-to-day engineering, you are affected until Anthropic clears the incident. The practical moves are to confirm scope on the official status page, pause autonomous agent jobs that will retry destructive commands, and fall back to local git workflows rather than routing production secrets through a substitute chat tool.

The outage lands on top of a still-open Microsoft 365 connector incident Anthropic first posted November 12, 2026. Mail, calendar, and search tools that route through that connector were already failing for most requests. Coding agents that reach Outlook or Teams through Claude integrations therefore hit two separate upstream failures at once.

For news background on Microsoft-side service disruption this week, explainx.ai covered the Microsoft 365 Exchange Online incident EX1550543 as a separate timeline. Beam focuses on what Anthropic published and what that means on developer laptops.

TL;DR

What Anthropic published on November 13, 2026

Anthropic's status page incident titled "Elevated errors on claude.ai, Claude Code, and platform.claude.com" opened at 09:58 UTC on November 13, 2026. The first line reads: "We are investigating an issue with users authenticating to Claude.ai and Claude Code."

On the same page, component summaries showed:

• claude.ai — Major Outage
• Claude Code — Partial Outage
• Claude Console (platform.claude.com) — Partial Outage
• Claude Cowork — Degraded Performance
• Claude API (api.anthropic.com) — Operational

That split matters for teams that route production traffic through the HTTP API while engineers use Claude Code interactively. An API-only integration might keep running while laptops cannot open a coding session—a pattern that still breaks human-in-the-loop review if your release process assumes an agent on the desktop.

Anthropic had not linked this November 13 incident to the earlier Microsoft 365 connector problem in the text we fetched. Treat them as parallel incidents until Anthropic states otherwise.

Update (November 13, 2026): Microsoft root cause for M365 connectivity

While Anthropic worked on claude.ai authentication, Microsoft published a recovery and root-cause statement for the broader Microsoft 365 connectivity outage that had already broken Claude's Microsoft 365 connector since November 12. Reporting that quoted Microsoft's X post—including The Washington Post and ABC News—attributes these lines to the company:

• "We are deeply sorry for the disruption this incident has caused our customers and partners."
• "We are seeing broad recovery and stability across our services."
• "This was a technical issue not related to cyber or AI / agents."

Microsoft described the failure as an issue introduced during an earlier network capacity upgrade that surfaced after a more recent network change, and said it had addressed the underlying issue while restoring remaining regions. Press summaries tied the disruption to incidents tracked as MO1550821 (Microsoft 365 connectivity) alongside Exchange Online work such as EX1550543—both referenced on status.cloud.microsoft.

For coding-agent users, three details matter:

1. Microsoft explicitly ruled out cyber and "AI / agents" as the cause in the statement reporters quoted. That does not clear your connector tokens or MCP configs, but it means you should not treat the event as a prompt-injection or agent-abuse story without separate evidence.
2. Copilot and Graph-dependent tools sat in the same outage footprint as Outlook and Teams in press reporting. Agents that call mail, calendar, or search through Microsoft's cloud should stay disabled or in degraded mode until both status.cloud.microsoft and Anthropic's connector line are green.
3. Staged recovery means retry storms hurt. Microsoft told outlets it restored in stages to preserve stability; autonomous agent loops that hammer Graph should back off the same way we advise in the November 12 explainx timeline.

Nothing in Microsoft's quoted statement changes Anthropic's separate authentication incident on November 13. Continue to treat login failures and M365 tool errors as two monitors, not one root cause.

The stacked Microsoft 365 connector failure

Starting November 12, 2026, Anthropic posted "Elevated errors affecting the Microsoft 365 connector." Updates through 22:23 UTC said the issue is caused by a Microsoft 365 service incident, that mail, calendar, and search tools in the connector still fail for most requests, and that Claude for Outlook remains affected. Anthropic pointed operators to status.cloud.microsoft and said it would update when service restores.

For agent users, the mechanism is straightforward:

1. Your coding agent (Claude Code or Cowork) holds OAuth or connector tokens for Microsoft 365.
2. Tool calls that list mail, search SharePoint, or read calendar entries go through Anthropic's connector layer.
3. Microsoft's upstream outage returns errors even when Claude's core model endpoint is healthy.
4. Autonomous loops that retry tool calls may hammer failing APIs, burn context, or—worse—fall back to shell commands the operator did not intend.

If your workflow uses Claude as a mail-aware coding agent, assume read/write mail tools are unsafe to rely on until both status pages clear. Our MCP security guide applies the same discipline to any connector whose remote side is unhealthy: disable the tool, do not let the model improvise around it.

Attack chain and failure modes (what actually breaks)

This is not a CVE-style exploit chain. It is an availability and trust-boundary sequence that still changes how much you should let an agent act without you:

1. Auth dependency. Claude Code normally proves identity against Anthropic's cloud. When authentication fails, the CLI may ask for /login, hang on the browser step, or log you out mid-conversation—behaviors HN users described on the Tell HN thread.
2. Session loss. Partial outages can strand half-finished tasks: the model stops responding, but local files may already be modified. That is the same recovery problem as an agent crash, without malware involved.
3. Retry amplification. CI jobs or long-running agent scripts that restart on error may open multiple login tabs, duplicate worktrees, or re-run expensive commands once auth flickers back.
4. Connector coupling. M365 tool failures look like agent "confusion"—the model may suggest workarounds (export mail to disk, run curl, paste tokens) unless you pause tools explicitly.
5. Vendor switching under pressure. Outages push engineers to paste logs into other chat products. That is a common credential exfiltration path even when every vendor is honest—accidental paste, not prompt injection.

Recent agent incidents we have covered—malicious AGENTS.md hijacking, GitBook rotating tokens after agent-tagged abuse, and Northflank's control-plane secret rotation after AI key abuse—all assume the agent can still reach its API. Today's failure mode is different: the control plane never comes online.

What developers are saying

Hacker News users reporting from multiple regions confirmed the failure matched Anthropic's status components:

• _jx reported confirmation from Italy and noted status.claude.com showed Major Outage at the time of the comment.
• Cyuonut said login worked briefly, then the CLI asked for /login and the browser flow hung on the claude.ai side.
• t0duf0du described repeated logouts mid-conversation and multiple login attempts failing.
• masteranza linked to status.claude.com and attributed the M365 connector line to Microsoft servers, matching Anthropic's November 12 updates.
• JanSt reported the service was down ("Yes its done").

These are anecdotal availability reports, not a reproduction guide. Use them to calibrate expectations, not as a substitute for the status page.

Recent Anthropic availability context

The same status history shows shorter incidents in the days before November 13:

• November 10, 2026 — Elevated errors on Fable 52 and Mythos 52 for about four minutes, marked resolved.
• November 8, 2026 — Degraded performance for Claude Fable 5.1, later resolved across claude.ai, the API, Claude Code, and Cowork with a published window from 00:18–00:32 PT.

Those were model-endpoint blips. The November 13 incident is authentication-wide on consumer surfaces, which blocks starting sessions entirely—a higher-severity class for interactive coding.

The Paris Call for AI Safety and Security that labs signed on November 9 asks for fail-safe shutdowns and faster serious incident sharing. Today's outage is operational rather than a deliberate model pause, but it is a live test of whether your team has a local fallback when a vendor control plane disappears.

What this means if you run AI coding agents

Confirm scope before you change architecture. Refresh status.claude.com and, if you use M365 tools, status.cloud.microsoft. If only Claude Code is red but your API monitors are green, your automation may still be running while humans cannot supervise—decide whether that gap is acceptable.

Stop unattended jobs first. Pipelines that invoke Claude Code headlessly should fail closed for the duration: do not let a wrapper script retry rm, git push, or infrastructure applies on backoff. The Recorded Future roundup of AI incidents in coding tools is a useful reminder that most damage happens when nobody is watching.

Record versions you were on. When service returns, note claude --version and any extension versions (see our Postman installer hook post for why pinned versions matter after disruptive weeks). If Anthropic publishes a post-mortem, you will want that baseline.

Work locally without outsourcing secrets. Commits, tests, and code review in git do not require Claude. If you must use another model temporarily, redact env files, customer data, and internal URLs—the same discipline as sharing Claude Code permission settings on public GitHub.

Disable broken MCP and connector tools. Turn off Microsoft 365, mail, or other flaky integrations in Claude Code settings until both vendors green-light their components. A model denied its calendar tool is safer than one guessing how to fetch mail via shell.

Watch for phishing noise. Outages attract fake status pages and typosquats. Our TypoPrompting typosquat write-up documents how claude.ai lookalikes profit from rushed typing. Bookmark the real status URL; do not click "fix Claude" ads.

What Beam can and cannot do here

Beam is local observability and hook-based policy for supported coding agents—it does not host Claude or repair Anthropic authentication.

What still works on your machine:

• Local git and filesystem work unchanged by the cloud outage.
• If Claude Code does start and hooks fire, Beam's collector can record and redact tool events under workspace policy, as described in agent hooks documentation.
• With workspace enforce enabled, PreToolUse deny for Claude Code and Codex can still block matching risky commands when the session is actually running—heuristic, hook-based, and fail-open if cached policy is missing, per troubleshooting guidance.

What does not work:

• Beam cannot log in for you, proxy Anthropic traffic, or keep Claude Code alive when the CLI never authenticates.
• Beam does not monitor claude.ai web sessions, Dispatch, or browser extensions—see our ThisBash Dispatch report for why that boundary matters on different days.
• Beam does not validate whether Microsoft's connector recovery succeeded.

Use the outage to review what your agents could have done last week: open what Beam watches, export recent events after service returns, and tighten enforce rules before the next incident—cloud or supply-chain.

Frequently asked questions

What is failing during the November 13, 2026 Claude outage?

Anthropic status.claude.com lists a major outage on claude.ai and partial outages on Claude Code and platform.claude.com, opened at 09:58 UTC with users unable to authenticate. Claude Cowork shows degraded performance while the Claude API remained operational on the same status snapshot.

Is the Claude API down too?

On the November 13 status snapshot Anthropic published, api.anthropic.com remained operational while claude.ai, Claude Code, and the console showed elevated errors. Treat that as a point-in-time reading and refresh the status page before you change production API routing.

Why are Claude Microsoft 365 mail and calendar tools still broken?

A separate incident that started November 12 ties the Microsoft 365 connector to an upstream Microsoft service problem. Anthropic says mail, calendar, and search tools fail for most requests and points customers to status.cloud.microsoft for Microsoft progress.

Can I keep using Claude Code offline during the outage?

Authentication to Anthropic must succeed for normal cloud-backed Claude Code use. Local files and git history on your machine are unaffected, but new sessions that require claude.ai login will fail until Anthropic resolves the incident.

Should I paste server fix prompts into another model while Claude is down?

Only if you strip secrets first. Outages push teams toward alternate chat tools quickly; that multiplies the risk of pasting production logs or API keys into a second vendor. Redact before you paste, and prefer local git operations until your primary agent works again.

Does Beam keep recording events when Claude Code cannot log in?

Beam records hook events for supported agents when the CLI can run and reach the local collector. It does not restore Anthropic authentication, does not proxy models, and does not guarantee capture if the Claude Code process never starts a session.

Related reading

• Paris Call for AI Safety: shutdown pledges and coding-agent expectations
• Malicious AGENTS.md in Downloads hijacking Codex and Claude Code
• GitBook incident: rotate API tokens after agent-tagged abuse
• Northflank secret rotation after AI key abuse
• MCP security: tool poisoning and connector risk
• Recorded Future on incidents involving coding tools
• Beam: how monitoring works and local-first storage

Incident details follow Anthropic's status.claude.com posts as of November 13, 2026 UTC. Components and mitigations can change without notice; refresh the status page before acting on production systems. Microsoft connector status is reported by Anthropic as upstream-dependent.