When AWS’s Middle East region misfired in March 2026, a misconfigured routing update turned into cascading failures across three availability zones in twelve minutes flat. Claude went offline for millions of users.
Banking and fintech apps across Argentina and Brazil started throwing authentication errors. None of those organisations had a direct contract with the network engineer who pushed the bad config.
That is the risk this article is actually about: the AI and infrastructure you never contracted with, sitting two or three layers inside a vendor you did.
The 30-Second Take
Three real 2026 incidents show the same pattern.
AWS’s March outage cascaded through 84-plus dependent services in minutes.
Microsoft Copilot’s five-hour outage on 1 June 2026 generated more than 14,000 incident reports, with 72% of users unable to load the panel at all.
Anthropic’s Claude experienced a cascading failure in March where login issues triggered problems in Opus 4.6, then Haiku 4.5 — and organisations without a multi-model failover strategy lost AI functionality entirely.
None of these were caused by the affected companies. All of them were felt by the affected companies’ customers.
Regulatory Net Is Closing Around This Exact Gap
From 2 August 2026, the EU AI Act’s obligations for general-purpose AI model providers become fully enforceable, fines included.
Providers must give downstream integrators enough information to understand a model’s capabilities and limits; deployers, in turn, must monitor the system’s operation and tell the provider and regulators immediately if a risk shows up.
That is a direct, regulatory answer to the AWS and Copilot pattern: you don’t get to say the failure was three layers removed from you if you never built the monitoring to see it coming.
This Is Now a Named Risk Category, Not a Hypothetical
Analysts are no longer treating this as an edge case. Gartner’s supply-chain risk research now names it directly: a “fourth-party” model provider you never contracted with may sit inside a vendor’s product, and a breach or outage at that fourth or fifth-party provider can cascade through the chain to hit you.
If your vendor risk assessment stops at the vendor you signed with, you are assessing the wrong layer.
Questions Your Board Should Be Asking Now
- Can we map which of our critical vendors have AI embedded in their platform that we did not directly select or contract?
- If our AI or cloud vendor’s vendor had an outage tomorrow — AWS-in-March-2026 style — do we know which of our services would go down, and for how long?
- Do we have a multi-model or multi-vendor failover strategy for any AI-dependent process, or does one provider’s outage mean total loss of function?
- Are we ready to meet EU AI Act deployer obligations — monitoring operation and reporting risk immediately — for every GPAI-based system we use, not just the ones we built ourselves?
- Who owns the answer when a fourth-party AI failure disrupts a customer-facing service: the vendor relationship owner, IT, or risk?
- When did we last test resilience against a cascading vendor failure rather than a single point of failure?
Keeping The Board Ahead Of The Curve
The Innovation of Risk tracks exactly this shift — from single-vendor oversight to nested AI supply-chain risk. If your board wants a structured way to map AI exposure and test your resilience against the next cascading outage, that is a practical place to start.