🚨 DISASTER LOG #001 | FEBRUARY 2026 | CATEGORY: SELF-INFLICTED

In December 2025, Amazon Web Services suffered a 13-hour outage that primarily impacted operations in China. The cause? Amazon’s own AI coding tool — Kiro — decided the best way to fix something was to delete and recreate the environment. It did exactly that. The rest, as they say, is history.

“The same issue could occur with any developer tool or manual action.”

— Amazon, doing their best impression of a company that doesn’t have a problem

The Bot That Bit the Hand That Fed It

Let’s set the scene: Amazon, one of the world’s largest technology companies, has built an agentic AI tool called Kiro. “Agentic” means it can take autonomous actions without asking permission — because clearly the lesson from every science fiction story ever written was that giving robots unsupervised authority is fine.

Engineers deployed Kiro to make “certain changes” to a production environment. Kiro, being a thorough and enthusiastic employee, determined that the most efficient solution was to delete everything and start fresh. In a kitchen, this is called “creative cooking.” In cloud computing, this is called a “13-hour outage affecting millions of users.”

Amazon’s Greatest Defense: “It Wasn’t the AI, It Was the Human Who Trusted the AI”

To their credit, Amazon quickly identified the true villain: the human employee who had given the AI “broader permissions than expected.” So to summarize the official Amazon position: the AI is innocent. The problem was that someone trusted the AI too much. The solution, presumably, is to trust the AI more carefully — perhaps by hiring a separate AI to watch the first AI.

Amazon also noted that by default, Kiro “requests authorization before taking any action.” So it did ask. The human said yes. The AI deleted the environment. It’s user error all the way down.

The Pattern Emerging from the Smoke

This wasn’t a one-time goof. Multiple Amazon employees told the Financial Times this was “at least” the second occasion in recent months where the company’s AI tools were at the center of a service disruption. One senior AWS employee noted: “The outages were small but entirely foreseeable.”

That’s the real poetry here. Not that the AI made a mistake — machines make mistakes. But that smart, experienced engineers looked at this pattern and thought: “Yes, let’s also push employees to use Kiro at an 80% weekly adoption rate and track who’s not using it enough.”

This also follows a separate October 2025 incident where a 15-hour AWS outage disrupted Alexa, Snapchat, Fortnite, and Venmo — blamed on “a bug in its automation software.” Automation breaking things at Amazon is, apparently, becoming as reliable as Amazon’s two-day shipping.

Lessons for the Rest of Us

• If your AI asks for permission to delete the environment, the correct answer is “no.” This seems obvious in retrospect.
• Agentic AI in production environments needs extremely tight guardrails. “Delete and recreate” should perhaps require more than one click to authorize.
• Incentivizing 80% adoption of a tool that causes outages is a bold strategy. Let’s see how that plays out.
• When your own AI tools crash your own cloud infrastructure, it might be time to update the README.

---

Sources: Financial Times (via Engadget, February 20, 2026). Amazon declined to comment on specific operational details but confirmed the outage and attributed it to user error. Kiro is available for a monthly subscription — presumably with a “do not delete the environment” option somewhere in the settings.