Outages & Downtime

AWS Outage Exposes the Danger of Agentic AI: When Amazon's Own Bot Took Down the Cloud

· 3 min read
Cloud computing infrastructure with server racks

Amazon’s AI Coding Assistant Caused a 13-Hour Production Outage

Here’s a bitter irony for the AI age: Amazon Web Services, the backbone of half the internet, was taken down by its own AI.

In December 2025, AWS suffered a 13-hour outage that knocked out the AWS Cost Explorer service in China. The culprit? Kiro, Amazon’s own AI coding assistant. According to reports from the Financial Times and Capacity Media, engineers allowed the agentic AI tool to make autonomous changes to a production system. Kiro determined the best course of action was to delete and rebuild the entire environment. It was wrong.

The Corporate Response: Blame the Humans

Amazon’s official response? User error. An AWS spokesperson told Capacity: “This brief event was the result of user error—specifically misconfigured access controls—not AI.”

Sure. The tool that Amazon designed to autonomously modify production systems did exactly what it was built to do—make decisions without human oversight—and when it nuked a live environment, it’s somehow the humans’ fault for trusting it.

This wasn’t Kiro’s only incident. Reports indicate multiple AWS outages in December were linked to errors involving the company’s own AI tools. After a major October outage that took down Reddit, Snapchat, Lloyds Bank, and Venmo, you’d think Amazon would pump the brakes on autonomous AI agents touching critical infrastructure. Nope.

The Bigger Picture: AI That Can’t Actually Work

The AWS disaster isn’t an isolated incident—it’s a symptom of a larger problem. A new study published in February 2026 called the Remote Labor Index put frontier AI models to the ultimate test: real paid freelance work.

The results were brutal. Even the best-performing model (Claude Opus 4.5) failed 96.25% of tasks. Not because AI is useless, but because it excels at drafts while completely failing at professional delivery—file formats, consistency, following briefs, all the boring details clients actually pay for.

Why This Matters: The Agentic AI Gamble

Here’s what should terrify you: companies like Amazon, Anthropic, and OpenAI are betting everything on “agentic AI”—systems that take actions with minimal human input. OpenAI’s Sam Altman recently said at the Cisco AI Summit: “Companies that are not set up to quickly adopt AI workers will be at a huge disadvantage.”

But what happens when those AI workers delete your production environment because they think it’s the right call?

The AWS outage is a warning shot. We’re handing autonomous decision-making power to systems that fail 96% of the time at basic tasks, then acting surprised when they break things. Amazon implemented “additional safeguards” and “mandatory peer review” after the fact—but only after a 13-hour outage proved their AI couldn’t be trusted.

The hype says AI will run our infrastructure. The reality? It can’t even run itself.

Sources: