Five Days, Four Platforms: Correlated Outages and AI Agent Concentration

Authors: Cloud Security Alliance AI Safety Initiative
Published: 2026-11-19

Categories: AI Resilience and Concentration Risk
Download PDF
Download Markdown

Key Takeaways

Between November 12 and November 15, 2026, Microsoft 365, Anthropic, OpenAI, and Amazon Web Services all reported service disruptions, and Microsoft’s preliminary post-incident review followed on November 16. These four platforms are among the most widely used foundations for enterprise AI agent deployments. Taken one at a time, each incident had a bounded scope and an explanation, either stated or pending. Taken together, they show how agent architectures concentrate risk in places that model-selection decisions do not reach. Five lessons stand out.

Model diversity did not buy resilience at the data layer. A Microsoft 365 networking gateway failure broke Microsoft 365 Copilot and Anthropic’s Microsoft 365 connector, and OpenAI reported errors on its own Microsoft connector in the same window [1][3][4][5]. An enterprise that hedged between two or three model vendors for its email and document agents would likely have seen all of them degrade at the same time.

Identity and session control planes are single points of failure that often go unseen. Anthropic’s November 13 sign-in incident disrupted authentication to claude.ai and Claude Code, while the Claude API was not listed as affected [6]. AWS’s November 15 us-east-1 incident disrupted IAM Identity Center sign-in [8][9]. Because Identity Center stores its configuration in the region where it was enabled, customers whose workloads run elsewhere can still be affected unless they have replicated the instance [10].

Agents fail partially and quietly. During Anthropic’s broader November 13 incident, scheduled tasks and routines ran more than two hours late and some were skipped [7]. Organizations that run unattended agent workloads should treat skipped or delayed runs as a reconciliation problem, not an availability footnote.

Resilience mechanisms spread the Microsoft failure. Microsoft’s preliminary review links the outage to a latent July defect, extra load from a Teams failover, and capacity already reduced for unrelated reasons. Its own failover and load-redirection mechanisms then shifted traffic onto fewer servers, triggering the latent defect across the fleet in under two hours [1][2].

Concentration should be measured by shared dependency, not vendor count. The useful question is how many independent control planes an agent workflow needs in order to complete, and which of them it shares with the fallback.

Background

This note builds on two earlier CSA research notes. The first examined how the Microsoft 365 outage disabled productivity assistants and security tooling together [14]. The second examined the broader November outage cluster, including the attribution problems that followed [15]. This note narrows the focus to one question: what the cluster reveals about AI agents running on shared control planes, and why multi-vendor model strategies did not deliver the independence many organizations assumed.

The largest event began on November 12, when Microsoft 365 suffered an outage of nearly 24 hours. BleepingComputer, reporting on Microsoft’s preliminary post-incident review, put the start at 11:20 AM ET on November 12 and the end of active customer impact at 11:10 AM ET on November 13 [1]; CIO, citing Microsoft’s service health dashboard, gave a start of 16:17 UTC (11:17 AM ET) [3]. Microsoft named three contributing factors. A software defect introduced in a July update to networking gateway servers had gone undetected. Microsoft Teams traffic had moved onto Exchange Online’s gateway infrastructure through a failover about 15 hours before the outage. And nearly 14% of gateway capacity was already offline because of an unrelated issue [1][2]. Charles Lamanna, Microsoft’s executive vice president for Copilot, Agents, and Platform, described “a cascading failure from server to server as the same amount of load was spread across fewer servers.” He added that “in less than two hours, the previously unknown defect was encountered across our fleet” [2]. At the low point, more than 80% of gateway servers were unresponsive [1][2]. CIO reported that the impact reached Exchange Online, Teams calendaring, Exchange-dependent SharePoint features, Microsoft 365 Copilot (which could lose access to email, calendars, and files), Defender for Office 365 Safe Links, parts of the Defender XDR portal, Intune, and Purview data loss prevention and labeling [3].

The effects reached other AI vendors almost immediately. Anthropic’s status page recorded elevated errors on its Microsoft 365 connector from 16:03 UTC on November 12 to 16:40 UTC on November 13 [4]. That start time, 11:03 AM ET, falls slightly before Microsoft’s reported start; the small gap likely reflects differences in how each provider timestamps detection versus declared impact. Anthropic’s initial update said requests to mail, SharePoint, Teams chat, and calendar were failing, later updates described the affected tools as mail, calendar, and search, and Claude for Outlook was also affected. Anthropic attributed the incident to “a Microsoft 365 service incident” at an upstream provider [4]. OpenAI posted its own incident, “Elevated errors affecting the Microsoft 365 connector,” warning that new connections might be broken and that tools were not working as expected. OpenAI did not state a cause. A third-party mirror of OpenAI’s status page records the incident as open from November 12 until it was resolved early on November 14 [5].

On November 13, Anthropic also had two incidents of its own, for which it has published no cause. From 09:58 to 11:58 UTC, users had trouble authenticating to claude.ai and Claude Code, with the Claude Console and Claude Cowork also listed as affected [6]. From 13:09 to 15:38 UTC, a wider incident degraded claude.ai, the Claude API, Claude Code, connectors, Claude in Meetings, Claude Managed Agents, and Claude Security. Anthropic’s 15:33 UTC update said that scheduled tasks and routines ran more than two hours late, with some skipped, and that adding new connectors was still unavailable [7]. Nothing in the public record links these two incidents to the Microsoft outage, and this note does not suggest a link.

On Sunday, November 15, AWS reported increased connectivity issues in us-east-1. Sign-ins, new account sign-ups, and some network connections returned errors for about three hours [8]. Contemporaneous reporting listed IAM Identity Center sign-in, AWS Sign Up, NAT Gateway, Elastic Load Balancing, Step Functions, CloudWatch, VPC Lattice, and Kiro, AWS’s AI development environment, among the affected services [8]; German trade coverage independently confirmed the Identity Center, Sign Up, and Kiro impacts [9]. AWS gave no cause in its closing notice [8]. The same period also brought a nationwide FAA ground stop and disruptions at JPMorgan Chase and United Airlines [12]. Those events fall outside the AI platform scope of this note.

The table below summarizes the AI-relevant incidents by the layer of the agent stack each one affected, separating causes the providers stated from those they did not.

Platform Date(s) (2026) Agent-stack layer affected Stated cause Agent-relevant impact
Microsoft 365 Nov 12–13 Data/tool plane (mail, calendar, files) and security controls Latent gateway defect, Teams failover load, reduced capacity (preliminary) [1][2] Copilot could lose access to mail, calendar, and files; Defender and Purview functions degraded [3]
Anthropic (Microsoft 365 connector) Nov 12–13 Third-party connector to Microsoft data plane Upstream Microsoft 365 incident [4] Claude connector requests (mail, calendar, SharePoint/search, Teams chat) and Claude for Outlook failed [4]
OpenAI (Microsoft connector) Nov 12–14 Third-party connector to Microsoft data plane Not stated; coincided with Microsoft incident [5] New connections could break; connector tools did not work as expected [5]
Anthropic (first-party) Nov 13 Authentication, then inference, agents, and scheduling Not published [6][7] Sign-in failures; Managed Agents degraded; scheduled tasks delayed or skipped [6][7]
AWS us-east-1 Nov 15 Identity and network control plane Not published [8] Identity Center sign-in errors; Kiro and networking services affected [8][9]

Security Analysis

The November cluster exposes four related patterns in how AI agents inherit risk from the platforms beneath them. The first concerns the limits of model-vendor diversity, the second the identity layer that precedes every model call, the third the way resilience mechanisms themselves propagate failure, and the fourth the quiet, partial failures that are specific to unattended agents. Together they point toward a different way of measuring concentration, which closes this section.

The Diversity Illusion

A common enterprise response to AI provider concentration risk has been to contract with more than one frontier model vendor and route requests through a gateway. CSA’s earlier work recommends this pattern [13], and it still protects against an outage at a single inference provider. The November 12 event, however, shows the pattern’s limits once an agent’s value comes from the tools and data it reaches rather than from the model alone. An email triage agent built on Microsoft 365 Copilot, a Claude agent using the Microsoft 365 connector, and a ChatGPT agent using OpenAI’s Microsoft connector all have different models and vendors. All three still read and write through the same Microsoft 365 service boundary, and the failed gateway sat in front of it [1][3][4][5]. Swapping the model does not swap the mailbox.

This failure mode appears structural, and it is likely to become more common as agents take on more work. As agentic deployments grow, an agent’s dependencies tend to shift from the model layer toward the tool layer: systems of record, collaboration suites, identity providers, ticketing platforms, and code hosts. Most of these are singular systems of record, since an organization cannot practically run its mailbox on two suites at once. The analytical consequence is that multi-model redundancy should be scored only against model-layer failures. Data-plane and tool-plane concentration should be assessed separately, against what an agent can still safely do when its primary system of record is unreachable.

Identity and Session Control Planes

The second pattern concerns the identity and session layer that comes before any model call. During Anthropic’s morning incident on November 13, interactive surfaces such as claude.ai and Claude Code failed at authentication, while the Claude API was not listed among the affected components [6]. A third-party analysis reported that Claude Code sessions showed Cloudflare “403 Attention Required” pages during that window, both at sign-in and mid-session. Anthropic has not confirmed any cause, and that observation should be treated as unverified [17]. This suggests that workloads authenticating with API keys could keep running, while workflows built on interactive sign-in, including developer agents in the terminal, could not. Organizations that had not recorded which authentication path each agent depended on would have found it hard to predict which agents would survive.

The AWS incident made the same point at the cloud layer. IAM Identity Center stores directory configuration, permission sets, and assignments in the region where it was enabled. Replication to additional regions is available, but the customer must configure it [10]. Dr. Web’s coverage noted that a company running workloads in Frankfurt with Identity Center configured in us-east-1 could still be locked out of sign-in [9]. Agents that get short-lived cloud credentials through federated human sessions, or through operator-initiated workflows that pass through Identity Center, inherit that home-region dependency. The October 2025 us-east-1 event, in which a DynamoDB DNS fault caused a long recovery, illustrated a similar coupling between us-east-1 control-plane services and authentication paths that customers in other regions relied on [11]. Identity is often among the most heavily shared and least redundant components in an agent’s dependency graph, and the November cluster suggests it deserves to be modeled as such.

Resilience Mechanisms as Propagation Paths

Microsoft’s preliminary explanation is notable because the components meant to provide resilience moved the failure around. A Teams failover concentrated load on Exchange gateway infrastructure. Health-based redirection then sent traffic from failing servers to healthy ones, which became overloaded and failed in turn [1][2]. Microsoft also reported that Teams, OneDrive, and SharePoint could recover sooner by switching to alternative infrastructure, while Exchange Online could not [1]. CSA’s November 18 note discusses this correlated-fragility pattern in detail [15]. For agent operators, the specific implication is that agent retry behavior forms part of the same feedback loop.

Agents differ from human users in how they respond to errors. A person who sees Outlook fail usually stops and tries later. An agent framework with default retry policies may keep issuing requests across many tasks, tenants, and scheduled runs. Anthropic’s status updates described delayed and skipped scheduled runs on its side [7], but no public evidence shows that agent traffic worsened the Microsoft incident, and this note makes no such claim. The underlying risk is still plausible enough to plan for. As agent populations grow, unbounded retries against a degraded shared dependency could add load just when that dependency is shedding it. Bounded retries, exponential backoff with jitter, and circuit breakers at the connector layer are therefore a shared responsibility, not just a local reliability choice.

Silent Partial Failure in Unattended Agents

The fourth pattern is the most specific to agents. Anthropic’s 15:33 UTC update said that scheduled tasks and routines ran more than two hours late and that some were skipped [7]. Late or skipped runs in an interactive chat product are a nuisance. In unattended agent workflows, such as nightly reconciliation, compliance evidence collection, security triage, or customer correspondence, a skipped run can leave a gap that no one notices until a downstream process depends on the missing output. A connector outage can also leave an agent running with a working model and broken tools. That agent may then reason from incomplete context, for example concluding that no new mail needs a response when the mailbox was simply unreachable, unless its tool layer separates “empty result” from “error.”

Microsoft’s incident also shows the security dimension of partial failure. CIO reported that Defender for Office 365 Safe Links might not have checked URLs in Outlook and Teams, and that Purview data loss prevention and automatic labeling were affected [3]. An agent that kept sending or summarizing content through a path that still worked while those controls were degraded would have operated with weaker guardrails than its risk assessment assumed. CSA’s November 14 note examined this co-failure between productivity and security tooling [14]. The issue can be more serious for agents, because they may keep acting at machine speed during the gap.

Measuring Concentration by Shared Dependency

Taken together, the four platforms suggest a different way to measure agent concentration. Counting vendors gives a misleading picture. A more useful measure is the set of independent control planes that a given agent workflow must reach to complete its task, covering identity, model inference, connectors, data stores, scheduling, and security enforcement. Each element of that set should be checked against the workflow’s designated fallback. If the fallback shares any element, the two paths are correlated, and the fallback gives no protection against failure of that element. The November 12 event is a clear example: the fallback model shared the data plane, so for Microsoft 365–dependent workflows the hedge offered no protection. By contrast, the same model-layer hedge would have helped during the November 13 Claude API degradation [7] for workloads that did not depend on Microsoft 365, which is precisely why diversity has to be scored layer by layer rather than in aggregate.

Recommendations

The recommendations below move from immediate clean-up of the November incidents, through engineering changes achievable within a quarter, to governance changes that alter how concentration risk is assessed and reported.

Immediate Actions

Organizations running AI agents against Microsoft 365, Anthropic, OpenAI, or AWS us-east-1 should first reconcile agent activity from November 12 through November 16. Agent and scheduler logs should be checked for runs that were skipped, delayed, or completed on empty or partial tool results. Any business process that consumed those outputs should be identified and, where needed, re-run. Particular attention belongs to agents that send external communications, because these may have acted while Safe Links, data loss prevention, or labeling controls were degraded [3]. Agents that make access decisions deserve similar scrutiny, because identity-plane degradation during the cluster may have left them working from failed lookups or incomplete directory data [6][8].

At the same time, identity teams should confirm the home region of every IAM Identity Center instance that agents or agent operators depend on. Where an instance sits in a single region, they should evaluate the multi-region replication AWS now documents [10]. Teams should also record which agents authenticate through interactive or federated sessions and which use workload credentials, since the November 13 incident indicated that these paths can fail independently [6].

Short-Term Mitigations

Over the next quarter, organizations should build a dependency map for each production agent workflow that covers five layers: identity, inference, connectors and tools, data stores, and scheduling. Each workflow’s fallback path should then be scored against the same map, with shared elements flagged as correlated dependencies. Where the data plane cannot be duplicated, as with a single collaboration suite, the realistic mitigation is a defined degraded mode rather than redundancy. In that mode the agent pauses, queues, or hands off to a person, instead of continuing on partial context.

The connector layer is where most of these failure modes become visible to the agent, so engineering teams should treat the controls in the following table as baseline requirements. Each addresses a specific behavior observed or plausible during the November cluster.

Baseline connector control Failure mode it addresses
Bounded retries with exponential backoff and jitter, plus circuit breakers that open on sustained upstream errors Retry storms adding load to a degraded shared dependency
Tool responses that distinguish “no results” from “dependency unavailable” Agents reading an outage as an empty mailbox or queue
Idempotency keys on write actions Retried or resumed runs duplicating messages, tickets, or transactions
Monitoring and alerts for missed or late scheduled runs, with automatic reconciliation when service returns Skipped or delayed unattended runs going unnoticed
Policy checks that halt sensitive agent actions when a dependent security control reports degraded status Agents acting with weaker guardrails than their risk assessment assumed

None of these controls is novel on its own. What the November cluster shows is that they need to be applied consistently at the connector boundary, where an agent’s model is still healthy but its view of the world is not.

Strategic Considerations

At the governance level, enterprises should update their AI concentration risk assessments so they measure shared control planes, not just vendor count, and report the results to the same risk committees that oversee cloud concentration. Third-party risk reviews of AI vendors should ask which upstream platforms each connector depends on, and how the vendor will show connector-level incidents separately from first-party ones. Anthropic’s and OpenAI’s separate status entries for their Microsoft 365 connectors [4][5] are useful models for that transparency. Contracts and service-level commitments for agent platforms should state how scheduled and unattended workloads are handled during an incident, including whether skipped runs are reported to the customer.

Finally, tabletop exercises should include a multi-platform scenario modeled on the November cluster. In that scenario the primary productivity suite, one model provider’s interactive sign-in, and a cloud identity home region all degrade within a few days of one another. Each event alone is manageable, but together they test whether degraded modes, reconciliation procedures, and fallback paths are truly independent. Financial entities subject to the EU Digital Operational Resilience Act may find that such exercises also support their ICT third-party concentration obligations [16].

CSA Resource Alignment

The most directly relevant prior CSA work is AI Provider Concentration Risk: Enterprise Resilience [13]. That paper describes AI provider dependency as a “kill switch” business continuity risk and introduces the concepts of API coupling, data coupling, and agentic coupling, along with the AI dependency inventory and degraded-mode fallback. The November cluster offers a field test of that framework. The multi-provider architecture it recommends addresses API coupling. The Microsoft 365 connector failures show that data coupling and agentic coupling can remain fully concentrated even after a model-layer hedge is in place. This note’s recommendation to score fallbacks against shared control planes extends that paper’s dependency inventory into a correlation analysis.

This note also follows two CSA research notes published during the outage cluster. When the Assistant and the Alarm Go Dark Together [14] examined how the Microsoft 365 outage took down productivity assistants and security tooling together, which grounds the security-control co-failure analysis above. The November Outage Cluster: Correlated Fragility and Attribution [15] covered the broader cluster, the role of failover and disaster recovery as propagation paths, and the attribution failures that followed. The present note narrows those findings to agent architectures and the gap between model diversity and resilience.

For control mapping, the AI Controls Matrix (AICM) v1.1 [18] provides the relevant objectives. These include its Business Continuity Management and Operational Resilience domain for degraded modes and dependency mapping, Supply Chain Management, Transparency, and Accountability for upstream connector dependencies, and Identity and Access Management for home-region and federation risks. MAESTRO [19] supports threat modeling of the cross-layer dependencies discussed here, especially failures in the deployment infrastructure and agent ecosystem layers that propagate upward into agent behavior.

References

[1] Abrams, Lawrence. “Microsoft says months-old bug caused global Microsoft 365 outage.” BleepingComputer, November 16, 2026.

[2] Kelly, Ross. “Microsoft 365 outage: Company exec lifts lid on the ‘cascading failure’ that caused lengthy disruption.” ITPro, November 16, 2026.

[3] Swain, Gyana. “Microsoft struggles to get M365 services back online after widespread outage.” CIO, November 13, 2026.

[4] Anthropic. “Elevated errors affecting the Microsoft 365 connector.” Claude Status, November 12–13, 2026.

[5] Pulsetic (mirror of OpenAI Status). “Elevated errors affecting the Microsoft 365 connector — OpenAI Incident #160452.” Pulsetic, November 2026.

[6] Anthropic. “Elevated errors on claude.ai, Claude Code, and platform.claude.com.” Claude Status, November 13, 2026.

[7] Anthropic. “Elevated errors on claude.ai, the Claude API, and Claude Code.” Claude Status, November 13, 2026.

[8] Pillitteri, Pasquale. “AWS down in US-EAST-1: sign-ins and sign-ups fail for three hours.” pasqualepillitteri.it, November 16, 2026.

[9] Seyfferth, Markus. “AWS-Störung in US-EAST-1: Identity Center betroffen.” Dr. Web, November 15, 2026 (updated November 18, 2026).

[10] Amazon Web Services. “IAM Identity Center Region data storage and operations.” AWS Documentation, accessed November 19, 2026.

[11] Amazon Web Services. “Summary of the Amazon DynamoDB Service Disruption in the Northern Virginia (US-EAST-1) Region.” AWS, October 2025.

[12] Das, Vidhya. “After ‘Black October’ comes ‘Blackout November’: A wave of global tech outages exposes the risks of growing digital dependency.” BusinessToday, November 17, 2026.

[13] Cloud Security Alliance. “AI Provider Concentration Risk: Enterprise Resilience.” CSA, June 2026.

[14] Cloud Security Alliance. “When the Assistant and the Alarm Go Dark Together: Concentration Lessons from the November 2026 Microsoft 365 Outage.” CSA, November 14, 2026.

[15] Cloud Security Alliance. “The November Outage Cluster: Correlated Fragility and Attribution.” CSA, November 18, 2026.

[16] European Insurance and Occupational Pensions Authority. “Digital Operational Resilience Act (DORA).” EIOPA, accessed November 19, 2026.

[17] Nazir, Tahir. “Claude went down twice on 13 November: what broke, what kept working, and what to do next time.” DEV Community, November 19, 2026.

[18] Cloud Security Alliance. “AI Controls Matrix (AICM) v1.1.” CSA, 2026.

[19] Cloud Security Alliance. “Agentic AI Threat Modeling Framework: MAESTRO.” CSA, February 6, 2025.

← Back to Research Index