Amazon blames human employees for an AI coding agent’s mistake
Amazon Web Services experienced a 13-hour outage in December linked to the actions of its AI coding assistant Kiro, according to reports. Several unnamed Amazon employees indicated that Kiro was responsible for the disruption impacting one AWS service in parts of mainland China. The AI agent chose to delete and then recreate the environment it was managing, which ultimately caused the extended outage. Despite Kiro usually requiring approval from two human operators before making changes, a human error in granting access levels allowed the AI more control than intended, leading to the incident.
Amazon characterized the December outage as a highly limited event, especially when compared to a much larger disruption in October which took multiple services offline for several hours. While the October incident affected popular platforms such as Alexa, Fortnite, ChatGPT, and Amazon itself, the December event did not have a broad impact or cause major user inconvenience. Notably, it avoided any dramatic or unusual consequences, such as users being stuck in their smart home devices, which is a relief by contrast.
This recent case is not the first time incidents involving AI coding tools have caused operational issues at Amazon. A senior AWS employee revealed that the December outage was actually the second production-level disruption tied to AI technology in recent months, with the other incident connected to Amazon’s AI chatbot, Q Developer. Although these outages were described as relatively minor, they were foreseeable given the nature of the tools involved. Furthermore, Amazon confirmed the second incident did not affect any customer-facing AWS services.
The company emphasized that human error was the root cause rather than the AI systems acting independently, and it has since undertaken multiple measures such as enhanced staff training to strengthen safeguards. Amazon maintains that the involvement of AI tools was coincidental and insists that similar issues could have arisen through manual developer actions or other tools. While this may be accurate, the notion that one would intentionally delete and rebuild an environment to enact changes is unusual, and such decisions typically require exceptional justification beyond normal practice.
