Amazon’s cloud division is facing scrutiny following reports that internal AI tools were involved in at least two service disruptions late last year. The incidents may have involved automated coding systems, and have raised questions about the guardrails needed when using AI agents in complex environments.
According to a report in the Financial Times, a mid-December event resulted in a roughly 13-hour interruption affecting a customer-facing system within Amazon Web Services. Engineers had enabled an AI coding assistant known as Kiro to implement changes in a production environment. The system, designed to act autonomously to assist developers, launched a sequence that removed and rebuilt part of the environment supporting a cost management service.
Amazon Said AI Was Not the Problem
Amazon disputes the characterization that AI was at fault. In a press statement, an AWS spokesperson said the disruption was due to improperly configured access controls and described the episode as limited in scope. The company maintained that core infrastructure services, including compute, storage, databases and AI offerings, were not affected. Only a single cost analysis tool in one region of China experienced an interruption, the company said.
A second event reportedly involved Amazon Q Developer, another AI-powered coding assistant. That incident, according to employees cited in reports, did not impact a customer-facing service. Amazon has emphasized that both situations were the result of human misconfiguration rather than AI or machine autonomy.
AI Systems Granted Permissions
Whatever the exact truth about the recent outages, reports indicate that prior to the December outage, AI systems within AWS had been granted permissions comparable to those of human IT staff. In typical production workflows, major changes usually require some form of secondary approval. Following the incidents, AWS implemented additional controls, including mandatory peer review for certain production-level actions and expanded staff training.
Beyond operational questions, the events have prompted concerns about AI governance within large enterprise environments. Amazon has been vocal about integrating generative AI into its development processes to improve efficiency. Chief executive Andy Jassy has previously described AI agents as tools that could shift employees away from repetitive tasks and toward more high value work.
At the same time, the company has announced significant workforce reductions over the past year, though leadership has framed those cuts as strategic rather than directly tied to automation.
Unpredictable Risks
As companies move to adopt AI, they face an unavoidable truth: AI-driven tools can introduce unpredictable risks. Unlike traditional scripts that require engineers to manually enter commands, autonomous agents can execute complex sequences exceptionally fast. That speed is a big advantage but it also compresses the window in which a human can prevent an error. AI systems operate within limited instructions and so lack awareness of downstream consequences, particularly in complex IT environments.
Amazon, for its part, has said there is no evidence that AI-enabled tools produce more mistakes than conventional development methods. The company has characterized the involvement of AI in the outages as coincidental and reiterated that developers must configure the scope of actions such tools can take.
