Amazon Web Services faced renewed scrutiny after reports linked a late 2025 service disruption to internal use of its AI coding assistant, Kiro. Although AWS maintains that human error caused the incident, the episode has intensified debate around how far companies should allow agentic AI to act inside production environments.
According to reports citing sources familiar with the matter, engineers permitted Kiro to make infrastructure changes that resulted in a 13 hour outage affecting AWS Cost Explorer in one of its Mainland China regions. Cost Explorer helps customers analyze and manage cloud spending, and while the disruption did not affect core compute, storage, or database services, it raised broader concerns about operational safeguards in AI assisted development workflows.
AWS disputes claims that Kiro independently triggered the downtime. In a statement, a company spokesperson said a misconfigured access control by an employee led to the issue, not the AI system itself. The company added that Kiro typically requires authorization before executing actions, and that an engineer had used a role with broader permissions than intended. AWS established extra security measures after the incident which included mandatory peer assessment for granting production access rights.
The agentic coding service Kiro enables users to create technical specifications and deployable code through natural language prompts. AWS positioned the tool as a way to streamline application development while avoiding earlier AI mishaps seen across the industry. However, reports suggest the tool opted to delete and recreate an environment during troubleshooting, which preceded the outage.
The event highlights a growing tension in cloud computing. On the one hand, AI coding assistants offer the promise of rapid development cycles and efficiency. On the other hand, they raise new governance issues, especially when engineers assign them high levels of permissions. Such incidents have also been reported elsewhere, such as when AI systems got stuck in a loop or used unexpected amounts of resources.
The AWS incident highlights the reality that as cloud service providers continue to integrate AI into their platforms, there is a need for clear access boundaries and human oversight, even when automation speeds up processes. In a complex cloud setup, teams can easily confuse the tool with the user.
