Amazon Cloud Unit Taken Down Twice By Its Own AI Tools: Report
Amazon’s cloud-computing arm suffered at least two recent service interruptions linked to the use of its own artificial intelligence coding assistants, prompting some internal concerns about the company’s rapid deployment of autonomous software agents inside production environments.
In mid-December, Amazon engineers allowed the company’s Kiro AI coding tool to implement system changes that ultimately led to a roughly 13-hour disruption affecting one of the systems customers use to analyze the cost of AWS services, people familiar with the matter told the Financial Times.
The agentic tool - which is capable of taking autonomous actions on behalf of users - reportedly determined that the optimal remediation step was to delete and recreate a computing environment. AWS later circulated an internal postmortem examining the outage.
Employees said the December incident marked the second time in recent months that one of Amazon’s internally deployed AI development tools had played a central role in a service disruption. In both cases, engineers permitted the software agent to execute changes without requiring secondary approval, a safeguard typically mandated for manual interventions in production systems.
AWS accounts for roughly 60% of Amazon’s operating profit and is investing heavily in artificial intelligence tools designed to function as independent “agents” capable of carrying out tasks based on high-level human instructions. The company - along with other large technology firms - is also positioning such tools for sale to external enterprise customers.
Amazon said it was a coincidence that AI tools were involved in the disruptions and maintained that the same outcome could have resulted from conventional development software or manual intervention.
“In both instances, this was user error, not AI error,” the company said, adding that it had found no evidence that mistakes occur more frequently when AI tools are involved.
The company described the December interruption as an “extremely limited event” affecting a single service in parts of mainland China and said the second incident did not impact a customer-facing AWS system.
Neither disruption approached the scale of a broader AWS outage in October 2025 that lasted approximately 15 hours and temporarily took multiple customers’ applications offline - including services operated by OpenAI.
Employees said the company’s AI development tools are often treated as operational extensions of human engineers and are granted comparable system permissions. In the December case, the engineer involved had broader access than anticipated - a user access-control issue that Amazon said allowed the changes to proceed without appropriate review.
AWS introduced Kiro in July as a next-generation coding assistant designed to go beyond so-called “vibe coding,” in which developers rapidly assemble applications using AI-generated suggestions. Instead, Kiro was intended to produce code directly from formal specifications.
Prior to Kiro’s launch, AWS engineers relied on Amazon Q Developer, an AI-powered chatbot designed to assist with software development. Employees said that tool was involved in an earlier outage.
Some staff members said they remain skeptical about the reliability of AI-assisted coding for mission-critical tasks, particularly as Amazon has set internal targets encouraging 80% of developers to use AI tools for coding at least once per week. The company is said to be closely monitoring adoption rates.
Amazon said customer uptake of Kiro has been strong and that it wants both clients and employees to benefit from efficiency gains. Following the December incident, AWS implemented additional safeguards, including mandatory peer review procedures and expanded staff training.
