Delegate system maintenance toAI Agent (Agentic AI) was used to do it, and it concluded that the fastest way to fix the bug was to "delete and rebuild" the entire environment? This sounds like a scene from a science fiction movie, but it has been accused of actually happening on Amazon's cloud servers.
Financial Times (recent)The whistleblower allegesLast December, AWS experienced a 13-hour service outage, the culprit being Amazon's own heavily promoted AI programming tools."Kiro"However, Amazon subsequently issued a statement strongly denying the allegations, emphasizing that it was entirely a case of "human error" and criticizing the report's content.There are many inaccuracies..
A builder shares why their workflow finally clicked.
Instead of jumping straight to code, the IDE pushed them to start with specs.
✔️ Clear requirements.
✔️ Acceptance criteria.
✔️ Traceable tasks.Their takeaway:
Think first. Code later.Get the full breakdown here 👉… pic.twitter.com/eD7ZrEdEn5
— Kiro (@kirodotdev) January 14, 2026
Financial Times: To fix a minor bug, AI decides to "delete and rebuild the entire environment".
According to the Financial Times, citing multiple sources familiar with the matter, the outage that occurred last December primarily affected AWS services in mainland China.
The incident stemmed from an engineer authorizing Kiro, Amazon's self-developed AI agent tool, to perform certain system changes. As a tool capable of autonomous action, Kiro, after evaluating the task, reached a rather "radical" conclusion: the best way to solve the problem was to delete and rebuild the entire environment.
This "decisive" decision ultimately led to a 13-hour service shutdown.
The report also revealed that this was "at least the second time" in recent months that the company's internal AI tools had caused a service outage, and stated that although the scale of such outages caused by allowing AI to handle problems on its own was not large, it was "completely expected."
Amazon issued a strong rebuttal: it was purely coincidental, and the mistake lay in "overstepping its authority."
In response to this highly damaging accusation, Amazon unusually released a full statement on its news blog to refute the allegations, characterizing the incident as a "user access control issue" rather than a loss of autonomy by the AI.
Amazon clarified several key facts in its statement:
• The impact is minimal:The outage affected only one service in a single region across 39 geographic regions worldwide—AWS Cost Explorer (a tool that helps customers visualize and manage AWS costs)—and did not impact core infrastructure such as computing, storage, databases, and AI technologies. No customer complaints were received during the outage.
• Excessive human intervention:By default, the Kiro tool requests authorization before performing any action. The problem was that the account used by the engineer in question had excessively high privileges, "beyond expectation." Amazon emphasizes that the same incorrect permission settings, whether using AI tools or manual operation, will lead to the same disaster.
• Denying a second disconnection:Amazon has strongly refuted the Financial Times’ claim of a “second AI-related outage,” calling it “completely false.”
A double-edged sword: When "AI agents" penetrate the core of enterprises
This Rashomon-like incident reflects the growing pains faced by tech giants as they actively adopt generative AI.
Since launching Kiro last July, Amazon has been strongly promoting its use among employees internally, even setting a high usage target of 80% per week. Compared to chatbots that only provide text suggestions in the past, "AI agents" like Kiro have the ability to actually write and execute code and modify systems.
While Amazon did experience a severe 15-hour outage last October affecting major clients like Alexa and Snapchat (at the time, the company attributed it to a bug in its automation software), this December incident highlights a different kind of practical risk: as AI begins to have execution capabilities, are traditional access control mechanisms rigorous enough?
Analysis of viewpoints
What the industry should be most wary of in this incident is not "whether AI will make mistakes", but "how much authority should we give AI".
Amazon's defense is logically sound: if engineers provide AI with a master key that can destroy the system, then the fault lies with the person who issued the key, not the AI that holds it. Kiro believes that "delete and rebuild" is the fastest solution, which may be correct in the pure logic of program operation, but it lacks human respect for the "online production environment" and an assessment of its commercial impact.
As more and more companies begin to experiment with enabling AI agents to perform operations with reduced human oversight, the AWS incident serves as a clear warning. Future cybersecurity and operations will shift their focus from simply "preventing external hackers" to "preventing internal AI assistants with excessive privileges." Like Amazon's subsequent mandatory "peer review" mechanism, this approach is not only indispensable in the AI era, but has become even more crucial than ever.
