If AI agents are going to touch real infrastructure, should the companies building them take responsibility when things break, or is “user error” a convenient escape hatch?
Amazon’s cloud division, Amazon Web Services (AWS), underwent at least two outages in December linked to its own AI tools, according to reports tied to Reuters and the Financial Times.
Here’s what happened.
In mid-December, a system AWS customers use to monitor their cloud costs was knocked offline for 13 hours. That wasn’t a typical hardware fault. It happened after engineers let an AI coding assistant named Kiro take action on its own. Instead of fixing a problem, the tool reportedly deleted and recreated the environment it was working on. And that broke the service.
That’s not just a glitch. It’s a scenario where an “agentic” AI with autonomy actually changed live infrastructure. And this wasn’t the only incident in recent months reportedly tied to AWS’s own AI tools.
Amazon insists the issue wasn’t the AI.
The company states the outage was user error tied to misconfigured access controls and would have happened with any developer tool, AI-powered or not. AWS also claims the second outage referenced in some reports didn’t occur inside AWS itself.
That response feels like damage control.
When your AI system can autonomously delete environments, that’s more than a simple misconfiguration. It raises real questions about checks and balances, permissions, and the autonomy these tools should have. Amazon’s stance that this was just a coincidence doesn’t fully address the bigger risk: when AI agents start making decisions without strict human oversight, small mistakes scale fast.
AWS is one of the most critical pieces of the internet’s backbone. It hosts countless services, apps, and business systems. If even a single cost-monitoring tool can go offline for over half a day because of an AI misstep, it shows the fragility of this AI-driven future.
There’s also a subtle tension here. AWS is pushing AI tools to developers and customers. At the same time, it wants to downplay risks when things go wrong. That contradiction matters.
