© Amazon
Putting AI in charge of a server may not be a very good idea. Amazon Web Services has experienced at least two service outages in the past few months. Cause: AI agents specialized in software development that caused outages by doing things at random. Amazon insists, however, that it was a human error.
The big players in AI have only one word on their lips: agents will do all the work for us. But before getting there — and causing a surge in unemployment —, some screws need to be tightened.
AI can make mistakes
The first incident, which occurred mid-December, caused about thirteen hours of downtime for a system allowing customers to explore AWS service costs. According to several internal sources cited by the Financial Times, engineers allowed the AI tool Kiro, capable of acting autonomously, to apply changes. The agent would have estimated that the best solution was to… « delete and recreate the environment », Oops. An internal report was then published internally.
Employees claim it is the second recent incident involving an AI tool from the company. In both cases, engineers would have let the agent fix a problem without additional human intervention. The outages would have been limited, but « completely predictable », according to an AWS executive.
When asked by the FT, Amazon denies any AI bug. It would actually be a « coincidence »: « the same problem could have occurred with any development tool or as a result of a manual action ». In this case, it happens that AI agents were in control, but « in both cases, it was a human error, not an AI error ». In this case, why were Kiro's safeguards reinforced after the December incident?
In the case of December, the engineer had access rights broader than intended, which would be an issue of authorization management — the AI agent wouldn't be involved in that. Amazon clarifies that Kiro requires default validation before any action. For the group, no element indicates that errors would be more frequent with AI tools. It's hard to imagine the company saying the opposite, given its massive involvement in the sector…
Amazon emphasizes that the December incident was an « extremely limited » event that affected only one service in certain regions of China. The second did not affect any customer-accessible AWS services. Neither of these two incidents had the scale of the AWS outage in October 2025, which caused a massive stir on the web worldwide.
Internally, employees would be skeptical about the reliability of these tools, which doesn't prevent the company from enforcing these solutions: the goal is that 80% of developers use AI at least once a week.
👉🏻 Follow tech news in real time: add 01net to your Google sources, and subscribe to our WhatsApp channel.
Source: FT
-
WhatsApp
